"""
2026.6.1
2026.6.1
5.5.0
0.24.0
__UNSLOTH_VERSIONING__
"""
# Unsloth auto generated code
# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
#
# This program is free software: you can redistribute it and/or modify
# it under the terms of the GNU Lesser General Public License as published by
# the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU Lesser General Public License
# along with this program. If not, see <https://www.gnu.org/licenses/>.
from torch import Tensor
import torch
import torch.nn as nn
from torch.nn import functional as F
from unsloth_zoo.temporary_patches.common import torch_compile
from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
from trl.trainer.grpo_trainer import (Any, AutoConfig, AutoModelForSequenceClassification, AutoProcessor, AutoTokenizer, BaseTrainer, DataLoader, Dataset, FSDP, GRPOConfig, GRPOTrainer, GenerationConfig, IterableDataset, Optional, Path, PeftConfig, PreTrainedModel, PreTrainedTokenizerBase, ProcessorMixin, RepeatSampler, RewardFunc, Sampler, SyncRefModelCallback, TrainerCallback, Union, VLLMClient, _ForwardRedirection, apply_chat_template, broadcast_object_list, datasets, defaultdict, deque, disable_dropout_in_model, ensure_master_addr_port, gather, gather_object, identity, inspect, is_conversational, is_datasets_available, is_flash_attn_2_available, is_liger_kernel_available, is_peft_model, is_rich_available, is_vllm_available, logger, logging, maybe_apply_chat_template, nanmax, nanmin, nanstd, nn, nullcontext, os, pad, partial, prepare_deepspeed, prepare_fsdp, prepare_multimodal_messages, print_prompt_completions_sample, profiling_context, profiling_decorator, seed_worker, selective_log_softmax, set_seed, shuffle_sequence_dict, split_pixel_values_by_grid, split_tensor_dict, textwrap, torch, transformers, unsplit_pixel_values_by_grid, unwrap_model_for_generation, AutoConfig, AutoModelForSequenceClassification, AutoProcessor, AutoTokenizer, Dataset, GRPOConfig, GRPOTrainer, GenerationConfig, IterableDataset, Optional, PeftConfig, PreTrainedModel, PreTrainedTokenizerBase, ProcessorMixin, RewardFunc, SyncRefModelCallback, TrainerCallback, Union, VLLMClient, datasets, defaultdict, deque, disable_dropout_in_model, ensure_master_addr_port, identity, inspect, is_liger_kernel_available, is_peft_model, is_vllm_available, logger, nn, os, pad, prepare_deepspeed, prepare_fsdp, set_seed, torch, transformers, Any, Union, gather, gather_object, is_conversational, logging, nanmax, nanmin, nanstd, os, pad, torch, FSDP, Optional, apply_chat_template, broadcast_object_list, gather, gather_object, is_flash_attn_2_available, maybe_apply_chat_template, nullcontext, os, pad, prepare_multimodal_messages, profiling_context, torch, transformers, unwrap_model_for_generation, os, pad, selective_log_softmax, torch, transformers, Any, Union, profiling_decorator, shuffle_sequence_dict, split_pixel_values_by_grid, split_tensor_dict, torch, unsplit_pixel_values_by_grid, PreTrainedModel, logger, os, torch, FSDP, nn, os, FSDP, nn, torch, GRPOTrainer, gather, inspect, nanmax, nanmin, os, pad, torch)
import os
import math
import logging
from typing import *
from dataclasses import dataclass, field
from packaging.version import Version
import torch
import numpy as np
from contextlib import nullcontext
from torch.nn import functional as F
import inspect
from transformers import DataCollatorForSeq2Seq, DataCollatorForLanguageModeling as TransformersDataCollatorForLanguageModeling
from transformers.training_args import ParallelMode
from unsloth_zoo.device_type import DEVICE_TYPE, device_synchronize
# Wrap trainer with padding to right and enable training mode
import functools
from types import MethodType
try:
from unsloth_zoo.gradient_checkpointing import reset_unsloth_gradient_checkpointing_buffers
except:
def reset_unsloth_gradient_checkpointing_buffers(): pass
def prepare_for_training_mode(f):
@functools.wraps(f)
def wrapper(self, *args, **kwargs):
# Finish the previous W&B run if this is a subsequent train() call.
# We do this at the START of train() (not the end) so that
# evaluate() / log() still work after train() completes.
# HF's WandbCallback.setup() will call wandb.init() for the new run.
# See: https://github.com/unslothai/unsloth/issues/3954
if getattr(self, '_unsloth_training_completed', False):
try:
import wandb
if wandb.run is not None:
wandb.finish()
# Reset HF's WandbCallback so it calls wandb.init() for the new run
for cb in self.callback_handler.callbacks:
if type(cb).__name__ == 'WandbCallback':
cb._initialized = False
break
except:
pass
# Enable training mode
_was_training = None
# Get gradient checkpointing setting from training arguments
use_gc = getattr(self.args, 'gradient_checkpointing', True)
if hasattr(self, 'model') and hasattr(self.model, "training"):
_was_training = self.model.training
if hasattr(self, 'model') and hasattr(self.model, "for_training"):
self.model.for_training(use_gradient_checkpointing=use_gc)
output = f(self, *args, **kwargs)
# Restore previous mode when possible
if hasattr(self, 'model') and hasattr(self.model, "for_inference"):
if _was_training is False:
self.model.for_inference()
elif _was_training is True and hasattr(self.model, "for_training"):
self.model.for_training(use_gradient_checkpointing=use_gc)
# Reset gradient checkpointing buffers to free memory while staying ready for next run
try:
reset_unsloth_gradient_checkpointing_buffers()
except:
pass
# Mark that training completed so the next train() call can
# finish this W&B run before starting a new one
self._unsloth_training_completed = True
return output
return wrapper
pass
torch_compile_options = {
"epilogue_fusion" : True,
"max_autotune" : False,
"shape_padding" : True,
"trace.enabled" : False,
"triton.enable_persistent_tma_matmul": torch.cuda.get_device_capability()[0] >= 9,
"cuda.cutlass_epilogue_fusion_enabled": torch.cuda.get_device_capability()[0] >= 9,
"cuda.cutlass_tma_only": torch.cuda.get_device_capability()[0] >= 9,
"cuda.compile_opt_level" : "-O2",
"cuda.enable_cuda_lto" : True,
}
@torch.compile(dynamic = True, fullgraph = True, options = torch_compile_options,)
def chunked_hidden_states_selective_log_softmax(
hidden_states: torch.Tensor,
lm_head: torch.Tensor,
index: torch.Tensor,
chunks: int = 4,
logit_scale_multiply: float = 0.0,
logit_scale_divide: float = 0.0,
logit_softcapping: float = 0.0,
temperature: float = 1.0,
) -> torch.Tensor:
# All Unsloth Zoo code licensed under AGPL3
flat_hidden_states = hidden_states.reshape(-1, hidden_states.shape[-1])
flat_index = index.reshape(-1)
chunked_hidden_states = torch.chunk(flat_hidden_states, chunks=chunks, dim=0)
chunked_index = torch.chunk(flat_index, chunks=chunks, dim=0)
all_per_token_logps = []
for chunk_hidden_states, chunk_index in zip(chunked_hidden_states, chunked_index):
chunk_logits = chunk_hidden_states.to(lm_head.dtype) @ lm_head.t()
if logit_scale_multiply != 0.0:
chunk_logits = chunk_logits * logit_scale_multiply
if logit_scale_divide != 0.0:
chunk_logits = chunk_logits / logit_scale_divide
if logit_softcapping != 0.0:
chunk_logits = logit_softcapping * torch.tanh(chunk_logits / logit_softcapping)
chunk_logits = chunk_logits.to(torch.float32)
if temperature != 1.0:
chunk_logits = chunk_logits / temperature
selected_logits = torch.gather(chunk_logits, dim=-1, index=chunk_index.unsqueeze(-1)).squeeze(-1)
logsumexp_values = torch.logsumexp(chunk_logits, dim=-1)
per_token_logps = selected_logits - logsumexp_values
all_per_token_logps.append(per_token_logps)
all_per_token_logps = torch.concat(all_per_token_logps)
all_per_token_logps = all_per_token_logps.reshape((hidden_states.shape[0], hidden_states.shape[1]))
return all_per_token_logps
@torch.compile(dynamic = True, fullgraph = True, options = torch_compile_options,)
def chunked_selective_log_softmax(
logits,
index,
temperature: float = 1.0,
chunks: int = 4,
):
chunked_logits = torch.chunk(logits.reshape(-1, logits.shape[-1]), chunks = chunks, dim = 0)
chunked_index = torch.chunk(index.reshape(-1), chunks = chunks, dim = 0)
all_per_token_logps = []
# Below loop does the same as selective_log_softmax(chunk_logits, chunk_index)
for chunk_logits, chunk_index in zip(chunked_logits, chunked_index):
chunk_logits = chunk_logits.to(torch.float32)
if temperature != 1.0:
chunk_logits = chunk_logits / temperature
selected_logits = torch.gather(chunk_logits, dim = -1, index = chunk_index.unsqueeze(-1)).squeeze(-1)
logsumexp_values = torch.logsumexp(chunk_logits, dim = -1)
per_token_logps = selected_logits - logsumexp_values
all_per_token_logps.append(per_token_logps)
pass
all_per_token_logps = torch.concat(all_per_token_logps)
all_per_token_logps = all_per_token_logps.reshape((logits.shape[0], logits.shape[1]))
return all_per_token_logps
def calculate_pad_tokens_in_prompt(
input_ids: torch.Tensor,
logits_to_keep: int,
pad_token_id: int
) -> torch.Tensor:
"""
Given prompt tensor, it returns all the left padded tokens in that sequence. so [pad, pad, pad, cat] = 3 tokens
"""
if logits_to_keep >= input_ids.shape[1]:
raise ValueError("logits_to_keep must be smaller than the sequence length.")
prompt_section = input_ids[:, :-logits_to_keep]
padding_mask = (prompt_section == pad_token_id)
pad_token_counts = padding_mask.sum(dim=1)
return pad_token_counts
def create_completion_attention_mask(
completion_input_ids: torch.Tensor,
left_pad_tokens_per_prompt: torch.Tensor,
max_left_pad: int,
pad_token_id: int
) -> torch.Tensor:
"""
Given that we have a sequence, [p,p,p,c,c,c,pad,pad,pad]
Where p are extra prompt tokens we got from slicing the torch tensor, c is completion tokens
and pad are pad tokens, this function would make a completion mask that would 0 out the pad
and p tokens. so in this example [0,0,0,1,1,1,0,0,0]
"""
batch_size, completion_len = completion_input_ids.shape
device = completion_input_ids.device
num_tokens_to_mask = max_left_pad - left_pad_tokens_per_prompt
indices = torch.arange(completion_len, device=device).unsqueeze(0)
shift_mask = indices >= num_tokens_to_mask.unsqueeze(1)
non_padding_mask = (completion_input_ids != pad_token_id)
final_mask = shift_mask & non_padding_mask
return final_mask
def left_pack_padding(tensor: torch.Tensor, pad_id: int) -> torch.Tensor:
"""
Moves all padding tokens in each sequence of a batch to the right.
"""
mask = (tensor != pad_id)
# Must do stable=True since binary mark is unordered
sorted_indices = torch.argsort(mask, dim=1, descending=True, stable=True)
packed_tensor = torch.gather(tensor, 1, sorted_indices)
return packed_tensor
def align_logprobs_with_mask(
logprob_tensor: torch.Tensor,
attention_mask: torch.Tensor,
pad_value: float = 0.0
) -> torch.Tensor:
"""
Aligns a log probability tensor with a given attention mask.
"""
device = logprob_tensor.device
batch_size, logprob_seq_len = logprob_tensor.shape
mask_seq_len = attention_mask.shape[1]
padded_logprobs = torch.full(
attention_mask.shape,
fill_value=pad_value,
dtype=logprob_tensor.dtype,
device=device
)
left_pad_counts = torch.argmax(attention_mask, dim=1)
cols = torch.arange(logprob_seq_len, device=device)
dest_indices = left_pad_counts.unsqueeze(1) + cols
# Create destination row indices
# Shape: [batch_size, logprob_seq_len]
row_indices = torch.arange(batch_size, device=device).unsqueeze(1).expand_as(dest_indices)
# --- 4. Filter out-of-bounds indices and perform assignment ---
# Create a mask to identify only the indices that are within the bounds
# of the target tensor's sequence length.
valid_mask = dest_indices < mask_seq_len
# Use this mask to select only the valid row indices, column indices,
# and the corresponding values from the logprob tensor.
# This flattens the selected elements into 1D tensors.
valid_rows = row_indices[valid_mask]
valid_cols = dest_indices[valid_mask]
valid_vals = logprob_tensor[valid_mask]
# Place the valid values into their correct positions in the padded tensor
# using a single, efficient advanced indexing operation.
padded_logprobs[valid_rows, valid_cols] = valid_vals
return padded_logprobs
def align_completion_tool_mask(
tool_mask: torch.Tensor,
completion_mask: torch.Tensor,
) -> torch.Tensor:
"""
Aligns a raw completion-length tool/env mask with Unsloth's repacked loss mask.
"""
if tool_mask is None:
return completion_mask
if tool_mask.shape[0] != completion_mask.shape[0]:
raise ValueError("tool_mask batch size must match completion_mask batch size.")
tool_mask = tool_mask.to(device=completion_mask.device)
if tool_mask.shape == completion_mask.shape:
aligned_tool_mask = tool_mask
else:
aligned_tool_mask = align_logprobs_with_mask(
tool_mask,
completion_mask,
pad_value=0,
)
return completion_mask * aligned_tool_mask.to(dtype=completion_mask.dtype)
def autotune_batch_and_chunks(
total_input_rows,
seq_len,
hidden_size,
vocab_size,
dtype_bytes=16,
multiplier=None
):
if multiplier is None:
final_m = max(4, seq_len // 4096)
else:
final_m = multiplier
if torch.cuda.is_available():
free_bytes, _ = torch.cuda.mem_get_info()
limit_gb = (free_bytes / (1024**3))*.80
elif hasattr(torch, "xpu") and torch.xpu.is_available():
# For XPU: estimate free memory from total - reserved
total_mem = torch.xpu.get_device_properties(0).total_memory
reserved_mem = torch.xpu.memory_reserved()
free_bytes = total_mem - reserved_mem
limit_gb = (free_bytes / (1024**3)) * 0.80
else:
# Fallback: assume 8GB available
limit_gb = 8.0
bytes_to_gb = 1024**3
b_vals = torch.arange(total_input_rows, 0, -1, device='cpu', dtype=torch.float32)
hidden_gb = (b_vals * seq_len * hidden_size * dtype_bytes) / bytes_to_gb
base_logits = ((b_vals/total_input_rows) * b_vals * seq_len * vocab_size * dtype_bytes) / bytes_to_gb
logits_gb = base_logits / final_m
total_mem_gb = hidden_gb + logits_gb
valid_mask = total_mem_gb <= limit_gb
valid_indices = torch.nonzero(valid_mask, as_tuple=False)
if valid_indices.shape[0] == 0:
#This means your GPU will OOM
return 4, final_m
best_idx = valid_indices[0].item()
final_b = int(b_vals[best_idx].item())
return final_b, final_m
def sanitize_logprob(logprob):
"""Local port of trl.scripts.vllm_serve.sanitize_logprob.
Filters NaN logprobs from vLLM outputs."""
value = logprob.logprob
if math.isnan(value):
logging.getLogger(__name__).warning(
f"Generated NaN logprob, token logprob '{logprob}' will be ignored"
)
return None
return value
def _unsloth_get_final_logit_softcapping(config):
"""Return final_logit_softcapping for a model config, falling back to the
nested text sub-config for composite models. Handles both:
- Gemma-4-style configs where the attribute lives on ``config.text_config``
- T5Gemma-style composite configs where the text sub-config is only
reachable via ``config.get_text_config()``
Returns 0 if unset, matching the previous behaviour.
"""
softcap = getattr(config, "final_logit_softcapping", None)
if softcap is None:
text_cfg = getattr(config, "text_config", None)
if text_cfg is None:
get_text_config = getattr(config, "get_text_config", None)
if callable(get_text_config):
try:
text_cfg = get_text_config()
except (TypeError, ValueError):
text_cfg = None
if text_cfg is not None and text_cfg is not config:
softcap = getattr(text_cfg, "final_logit_softcapping", None)
return 0 if softcap is None else softcap
def _unsloth_get_mm_token_id(processing_class, attr_name, token):
tokenizer = getattr(processing_class, "tokenizer", processing_class)
token_id = getattr(processing_class, attr_name, None)
if token_id is None:
token_id = getattr(tokenizer, attr_name, None)
convert_tokens_to_ids = getattr(tokenizer, "convert_tokens_to_ids", None)
if token_id is None and convert_tokens_to_ids is not None:
token_id = convert_tokens_to_ids(token)
if type(token_id) is int and token_id >= 0:
if token_id != getattr(tokenizer, "unk_token_id", None):
return token_id
return None
def _unsloth_fix_mm_token_type_ids(
processing_class, input_ids, mm_token_type_ids = None, completion_ids = None
):
image_token_id = _unsloth_get_mm_token_id(
processing_class, "image_token_id", "<|image_pad|>"
)
video_token_id = _unsloth_get_mm_token_id(
processing_class, "video_token_id", "<|video_pad|>"
)
if image_token_id is not None or video_token_id is not None:
rebuilt = input_ids.new_zeros(input_ids.shape)
if image_token_id is not None:
rebuilt = rebuilt.masked_fill(input_ids == image_token_id, 1)
if video_token_id is not None:
rebuilt = rebuilt.masked_fill(input_ids == video_token_id, 2)
return rebuilt
if (
mm_token_type_ids is not None
and completion_ids is not None
and mm_token_type_ids.shape[0] == input_ids.shape[0]
and mm_token_type_ids.shape[1] + completion_ids.shape[1] == input_ids.shape[1]
):
return torch.cat(
[mm_token_type_ids, mm_token_type_ids.new_zeros(completion_ids.shape)],
dim = 1,
)
return mm_token_type_ids
def _unsloth_clear_stateful_mrope(model):
modules = getattr(model, "modules", None)
if modules is None:
return False
cleared = False
for module in modules():
if hasattr(module, "compute_3d_position_ids") and hasattr(
module, "rope_deltas"
):
module.rope_deltas = None
cleared = True
return cleared
def grpo_compute_loss(
ref,
new,
old,
sampling_per_token_logps,
input_ids,
mask,
beta,
advantages,
**kwargs
):
# All Unsloth Zoo code licensed under AGPL3
# Set defaults for optional arguments
loss_type = kwargs.get("loss_type", "grpo")
epsilon_low = kwargs.get("epsilon_low", 0.2)
epsilon_high = kwargs.get("epsilon_high", 0.2)
max_completion_length = kwargs.get("max_completion_length", 8192)
delta = kwargs.get("delta", None)
importance_sampling_level = kwargs.get("importance_sampling_level", "token")
num_items_in_batch = kwargs.get("num_items_in_batch", None)
current_gradient_accumulation_steps = kwargs.get("current_gradient_accumulation_steps", 1)
num_processes = kwargs.get("num_processes", 1)
use_vllm = kwargs.get("use_vllm", False)
vllm_importance_sampling_cap = kwargs.get("vllm_importance_sampling_cap", 2.0)
get_sapo_token_loss = kwargs.get("get_sapo_token_loss", None)
sapo_temperature_pos = kwargs.get("sapo_temperature_pos", 1.0)
sapo_temperature_neg = kwargs.get("sapo_temperature_neg", 1.05)
get_gamma_weights = kwargs.get("get_gamma_weights", None)
vespo_k_pos = kwargs.get("vespo_k_pos", 2.0)
vespo_lambda_pos = kwargs.get("vespo_lambda_pos", 3.0)
vespo_k_neg = kwargs.get("vespo_k_neg", 3.0)
vespo_lambda_neg = kwargs.get("vespo_lambda_neg", 2.0)
get_off_policy_mask = kwargs.get("get_off_policy_mask", None)
off_policy_mask_threshold = kwargs.get("off_policy_mask_threshold", None)
input_ids = input_ids.unsqueeze(-1)
if advantages.dim() == 1:
advantages = advantages.unsqueeze(1)
if off_policy_mask_threshold is not None:
off_policy_mask = get_off_policy_mask(
advantages=advantages,
per_token_logps=new,
old_per_token_logps=old,
mask=mask,
off_policy_threshold=off_policy_mask_threshold,
)
with torch.no_grad():
if use_vllm and sampling_per_token_logps is not None:
#must filter out extra prompt tokens in begining after making input_ids left padded
importance_sampling_ratio = torch.exp((old * mask) - sampling_per_token_logps)
importance_sampling_ratio = torch.clamp(
importance_sampling_ratio, max=vllm_importance_sampling_cap
)
pass
# Must detach - otherwise gradients are not propagated correctly!
# exp(x - x) == 1
# loss_i = torch.exp(new - new.detach()) * advantages.unsqueeze(1)
if old is not None:
log_ratio = new - old
else:
log_ratio = new - new.detach()
if importance_sampling_level == "token":
log_importance_weights = log_ratio
elif importance_sampling_level == "sequence":
log_importance_weights = (log_ratio * mask).sum(-1) / mask.sum(-1).clamp(min=1.0)
log_importance_weights = log_importance_weights.unsqueeze(-1)
else:
raise ValueError(
f"Unknown importance sampling level: {importance_sampling_level}. Possible values are 'token' "
"and 'sequence'."
)
coef_1 = torch.exp(log_importance_weights)
# Reverse KL
# Note that this is a low variance low bias estimator for the KL divergence as used in GRPO paper
if beta != 0.0:
kl_i = torch.exp(ref - new) - (ref - new) - 1.0
else:
# set kl_i to a tensor of zeros with the correct shape
if importance_sampling_level == "sequence":
kl_i = new.new_zeros(new.size(0), 1)
else:
kl_i = torch.zeros_like(new)
# Full correct reverse KL divergence?? Missing term maybe?
# kl_i = torch.exp(new) * kl_i
# Below is forward KL (normal KL)
# kl_i = torch.exp(old) * (old - new)
if loss_type == "cispo":
clamped_ratios = torch.clamp(coef_1, max=epsilon_high).detach()
loss_i = -clamped_ratios * advantages * new
#breakpoint()
elif loss_type in ["grpo", "bnpo", "dr_grpo", "dapo"]:
coef_2 = torch.clamp(coef_1, 1 - epsilon_low, 1 + epsilon_high)
if delta is not None:
loss_1 = torch.clamp(coef_1, max=delta) * advantages
else:
loss_1 = coef_1 * advantages
pass
loss_2 = coef_2 * advantages
loss_i = -torch.min(loss_1, loss_2)
elif loss_type == "sapo":
if get_sapo_token_loss is None:
raise Exception(f"sapo is only available in TRL 0.26.0+")
loss_i = torch.empty_like(coef_1)
positive_advantages_mask = advantages.repeat([1, coef_1.shape[1]]) > 0
#since we have n_chunks some tensors may error if they dont have elements in them
if coef_1[positive_advantages_mask].numel() != 0:
loss_i[positive_advantages_mask] = get_sapo_token_loss(
coef_1[positive_advantages_mask], sapo_temperature_pos
)
if coef_1[~positive_advantages_mask].numel() != 0:
loss_i[~positive_advantages_mask] = get_sapo_token_loss(
coef_1[~positive_advantages_mask], sapo_temperature_neg
)
loss_i = -loss_i * advantages
elif loss_type == "vespo":
if get_gamma_weights is None:
raise Exception("vespo is only available in TRL 0.26.0+")
phi_seq = get_gamma_weights(
advantages=advantages,
log_ratio_per_token=log_ratio,
mask=mask,
importance_sampling_ratio=kwargs.get("importance_sampling_ratio"),
k_pos=vespo_k_pos,
lambda_pos=vespo_lambda_pos,
k_neg=vespo_k_neg,
lambda_neg=vespo_lambda_neg,
)
loss_i = -phi_seq * advantages * new
else:
raise ValueError(f"Unknown loss type: {loss_type}")
if off_policy_mask_threshold is not None:
loss_i = loss_i * off_policy_mask
if use_vllm and sampling_per_token_logps is not None:
loss_i = loss_i * importance_sampling_ratio
#delta for metric
with torch.no_grad():
delta = torch.abs(old - sampling_per_token_logps)
delta = delta * mask
flat_is_ratio = importance_sampling_ratio * mask
else:
delta = torch.tensor([]).detach()
flat_is_ratio = torch.tensor([]).detach()
if beta != 0.0:
loss_i = loss_i + beta * kl_i
mask = mask.to(torch.float32)
n_mask_per_reward = mask.sum(1)
# https://github.com/huggingface/trl/blob/e8b8499f1f8d76838155b515e414ee98f757d6d5/trl/trainer/grpo_trainer.py#L1624
if loss_type in ["grpo", "sapo"]:
loss = ((loss_i * mask).sum(-1) / mask.sum(-1).clamp(min=1.0)).mean()
loss = loss / current_gradient_accumulation_steps
elif loss_type == "bnpo":
loss = (loss_i * mask).sum() / mask.sum().clamp(min=1.0)
loss = loss / current_gradient_accumulation_steps
elif loss_type == "dr_grpo":
loss = (loss_i * mask).sum() / (loss_i.size(0) * max_completion_length)
loss = loss / current_gradient_accumulation_steps
elif loss_type in ["cispo", "dapo", "vespo"]:
normalizer = num_items_in_batch/ num_processes
loss = (loss_i * mask).sum() / normalizer
else:
raise ValueError(f"Unknown loss type: {loss_type}")
# loss = (loss_i * mask).sum() / mask.sum()
# Get metrics as well which are folded
def masked_batch_mean(x):
with torch.inference_mode():
completion_length = n_mask_per_reward.mean()
if x.shape[1] == 1: # when importance_sampling_level == "sequence"
return completion_length, x.mean()
else:
mean_kl_per_reward = (x * mask).sum(1) / n_mask_per_reward
mean_kl = mean_kl_per_reward.mean()
return completion_length, mean_kl
completion_length, mean_kl = masked_batch_mean(kl_i)
return loss, completion_length, mean_kl, delta, flat_is_ratio, coef_1, mask
class UnslothEfficientGRPO(torch.autograd.Function):
# All Unsloth Zoo code licensed under AGPL3
@staticmethod
def forward(ctx, _new_logps, _old_logps, _ref_logps, _sampling_per_token_logps, lm_head, _input_ids, _mask, _advantages, beta, scaler = None, n_chunks = 1, extra_kwargs=None):
if extra_kwargs is None:
extra_kwargs = {}
def compute_loss(new_logps, old_logps, ref_logps, sampling_per_token_logps, input_ids, mask, advantages, scaling):
loss, completion_length, mean_kl, delta, flat_is_ratio, coef_1, _mask = grpo_compute_loss(
ref_logps,
new_logps,
old_logps,
sampling_per_token_logps,
input_ids,
mask,
beta,
advantages,
**extra_kwargs,
)
# Scale loss if needed for mixed precision training
scaled_loss = loss * scaling
# Must add .loss.detach otherwise autograd uses 2x VRAM
return scaled_loss, (loss.detach(), completion_length, mean_kl, delta, flat_is_ratio, coef_1)
pass
device =_new_logps.device
grad_inputs = torch.empty_like(_new_logps)
accumulated_loss = torch.zeros(1, device = device)[0]
accumulated_completion_length = torch.zeros(1, device = device)[0]
accumulated_mean_kl = torch.zeros(1, device = device)[0]
accumulated_delta = []
accumulated_flat_is_ratio = []
accumulated_coef_1 = []
def accumulate_chunk(
new_logps_j,
old_logps_j,
ref_logps_j,
sampling_per_token_logps_j,
input_ids_j,
mask_j,
advantages_j,
scaling,
grad_inputs_j,
):
(chunk_grad_input,), (chunk_loss, (unscaled_loss, chunk_completion_length, chunk_mean_kl, chunk_delta, chunk_flat_is_ratio, chunk_coef_1)) = torch.func.grad_and_value(
compute_loss,
argnums = (0,),
has_aux = True,
)(new_logps_j, old_logps_j, ref_logps_j, sampling_per_token_logps_j, input_ids_j, mask_j, advantages_j, scaling)
accumulated_loss .add_(unscaled_loss)
accumulated_completion_length.add_(chunk_completion_length)
accumulated_mean_kl .add_(chunk_mean_kl)
accumulated_delta .append(chunk_delta)
accumulated_flat_is_ratio .append(chunk_flat_is_ratio)
accumulated_coef_1 .append(chunk_coef_1)
grad_inputs_j[:] = chunk_grad_input
pass
accumulate_chunk = torch.compile(
accumulate_chunk,
fullgraph = True,
# [TODO] Dynamic marking causes torch.compile errors if sequence length is long
dynamic = True,
options = torch_compile_options,
)
grad_inputs_chunks = torch.chunk(grad_inputs, chunks = n_chunks, dim = 0)
new_logps = torch.chunk(_new_logps, chunks = n_chunks, dim = 0)
if _old_logps is not None:
old_logps = torch.chunk(_old_logps, chunks = n_chunks, dim = 0)
else:
old_logps = [None] * n_chunks
if _ref_logps is not None:
ref_logps = torch.chunk(_ref_logps, chunks = n_chunks, dim = 0)
else:
ref_logps = [None] * n_chunks
if _sampling_per_token_logps is not None:
sampling_per_token_logps = torch.chunk(_sampling_per_token_logps, chunks = n_chunks, dim = 0)
else:
sampling_per_token_logps = [None] * n_chunks
input_ids = torch.chunk(_input_ids, chunks = n_chunks, dim = 0)
mask = torch.chunk(_mask, chunks = n_chunks, dim = 0)
advantages = torch.chunk(_advantages, chunks = n_chunks, dim = 0)
# Get mixed precision scaling if seen
scaling = scaler.get_scale() if scaler is not None else 1.0
# Force torch.compile to use dynamic shapes for seqlen dim
# mark_dynamic = lambda x: torch._dynamo.mark_dynamic(x, 1)
for (grad_inputs_j, new_logps_j, old_logps_j, ref_logps_j, sampling_per_token_logps_j, input_ids_j, mask_j, advantages_j, ) in\
zip(grad_inputs_chunks, new_logps, old_logps, ref_logps, sampling_per_token_logps, input_ids, mask, advantages):
# [TODO] Dynamic marking causes torch.compile errors if sequence length is long
# mark_dynamic(new_hidden_states_j)
# mark_dynamic(ref_hidden_states_j)
# if old_hidden_states_j is not None:
# mark_dynamic(old_hidden_states_j)
# mark_dynamic(input_ids_j)
# mark_dynamic(mask_j)
accumulate_chunk(
new_logps_j,
old_logps_j,
ref_logps_j,
sampling_per_token_logps_j,
input_ids_j,
mask_j,
advantages_j,
scaling,
grad_inputs_j,
)
pass
grad_inputs .div_(n_chunks)
accumulated_loss .div_(n_chunks)
accumulated_completion_length.div_(n_chunks)
accumulated_mean_kl .div_(n_chunks)
if _sampling_per_token_logps is not None:
accumulated_delta = torch.cat(accumulated_delta, dim=0)
accumulated_flat_is_ratio = torch.cat(accumulated_flat_is_ratio, dim=0)
else:
accumulated_delta = None
accumulated_flat_is_ratio = None
accumulated_coef_1 = torch.cat(accumulated_coef_1, dim=0)
ctx.save_for_backward(grad_inputs)
return (
accumulated_loss,
accumulated_completion_length,
accumulated_mean_kl,
accumulated_delta,
accumulated_flat_is_ratio,
accumulated_coef_1
)
pass
@staticmethod
def backward(ctx, grad_output, dcompletion_length, dmean_kl, ddelta, ddflat_is_ratio, dcoef_1):
(grad_input,) = ctx.saved_tensors
return (grad_input, None, None, None, None, None, None, None, None, None, None, None)
pass
def grpo_accumulated_loss(
trainer,
input_ids,
attention_mask,
logits_to_keep,
completion_mask,
advantages,
old_logps,
ref_logps,
n_chunks = -1,
tool_mask = None,
**kwargs,
):
# All Unsloth Zoo code licensed under AGPL3
bsz, qlen = input_ids.shape
pixel_values = kwargs.get('pixel_values',None)
image_grid_thw = kwargs.get('image_grid_thw',None)
pixel_attention_mask = kwargs.get('pixel_attention_mask',None)
image_sizes = kwargs.get('image_sizes',None)
num_images = kwargs.get('num_images',None)
# Transformers 5.x requires token_type_ids/mm_token_type_ids for some vision models
token_type_ids = kwargs.get('token_type_ids',None)
mm_token_type_ids = kwargs.get('mm_token_type_ids',None)
if mm_token_type_ids is not None or image_grid_thw is not None:
mm_token_type_ids = _unsloth_fix_mm_token_type_ids(
trainer.processing_class, input_ids, mm_token_type_ids
)
sampling_per_token_logps = kwargs.get("sampling_per_token_logps", None) if getattr(trainer, "vllm_importance_sampling_correction", False) else None
temperature = kwargs.get("temperature", 1.0)
logit_scale_multiply = kwargs.get("logit_scale_multiply", 0.0)
logit_scale_divide = kwargs.get("logit_scale_divide", 0.0)
logit_softcapping = kwargs.get("logit_softcapping", 0.0)
prev_max_left_pad = kwargs.get("max_left_pad", 0) #Always get max_left_pad for when training LLMs, enabled by deafult.
#Delete this from kwargs so less issues
_ = kwargs.pop("sampling_per_token_logps", None)
kwargs["vllm_importance_sampling_cap"] = trainer.vllm_importance_sampling_cap if sampling_per_token_logps is not None else None
kwargs["get_sapo_token_loss"] = trainer.get_sapo_token_loss if hasattr(trainer, "get_sapo_token_loss") else None
kwargs["sapo_temperature_pos"] = trainer.args.sapo_temperature_pos if hasattr(trainer.args, "sapo_temperature_pos") else None
kwargs["sapo_temperature_neg"] = trainer.args.sapo_temperature_neg if hasattr(trainer.args, "sapo_temperature_neg") else None
kwargs["get_gamma_weights"] = trainer.get_gamma_weights if hasattr(trainer, "get_gamma_weights") else None
kwargs["vespo_k_pos"] = trainer.args.vespo_k_pos if hasattr(trainer.args, "vespo_k_pos") else 2.0
kwargs["vespo_k_neg"] = trainer.args.vespo_k_neg if hasattr(trainer.args, "vespo_k_neg") else 3.0
kwargs["vespo_lambda_pos"] = trainer.args.vespo_lambda_pos if hasattr(trainer.args, "vespo_lambda_pos") else 3.0
kwargs["vespo_lambda_neg"] = trainer.args.vespo_lambda_neg if hasattr(trainer.args, "vespo_lambda_neg") else 2.0
kwargs["get_off_policy_mask"] = trainer.get_off_policy_mask if hasattr(trainer, "get_off_policy_mask") else None
kwargs["off_policy_mask_threshold"] = trainer.args.off_policy_mask_threshold if hasattr(trainer.args, "off_policy_mask_threshold") else None
kwargs["use_vllm"] = trainer.use_vllm
# Find closest multiple
factors = [i for i in range(1, bsz + 1) if bsz % i == 0]
if n_chunks == -1: n_chunks = bsz
n_chunks = factors[min(np.searchsorted(factors, n_chunks), len(factors)-1)]
if not hasattr(trainer, '_autocast_dtype'):
trainer._autocast_dtype = torch.float16 if os.environ.get('ACCELERATE_MIXED_PRECISION', 'fp16') == 'fp16' else torch.bfloat16
if os.environ.get('UNSLOTH_FORCE_FLOAT32', '0') == '1': trainer._autocast_dtype = None
pass
os.environ["UNSLOTH_RETURN_HIDDEN_STATES"] = "1"
lm_head = trainer.model.get_output_embeddings().weight
dtype_bytes = 16 if trainer._autocast_dtype in [torch.float16, torch.bfloat16] else 32
total_rows = input_ids.shape[0]
seq_len = input_ids.shape[1]
hidden_dim = lm_head.shape[1]
vocab_dim = lm_head.shape[0]
if trainer.args.unsloth_grpo_mini_batch is None:
if not hasattr(trainer, "_has_autotuned"):
trainer._has_autotuned = True
B, multiplier = autotune_batch_and_chunks(
total_rows, seq_len, hidden_dim, vocab_dim, dtype_bytes, trainer.args.unsloth_logit_chunk_multiplier
)
trainer.args.unsloth_grpo_mini_batch = max(1, total_rows//B)
trainer.args.unsloth_logit_chunk_multiplier = multiplier
B = trainer.args.unsloth_grpo_mini_batch
multiplier = trainer.args.unsloth_logit_chunk_multiplier
elif trainer._step % trainer.current_gradient_accumulation_steps == 0:
B = trainer.args.unsloth_grpo_mini_batch
multiplier = trainer.args.unsloth_logit_chunk_multiplier
del trainer._has_autotuned
del trainer.args.unsloth_grpo_mini_batch
del trainer.args.unsloth_logit_chunk_multiplier
else:
B = trainer.unsloth_grpo_mini_batch
multiplier = trainer.args.unsloth_logit_chunk_multiplier
else:
if trainer.args.unsloth_grpo_mini_batch > total_rows:
B = total_rows
else:
B = trainer.args.unsloth_grpo_mini_batch
if trainer.args.unsloth_logit_chunk_multiplier is None:
multiplier = max(4, seq_len // 4096)
else:
multiplier = trainer.args.unsloth_logit_chunk_multiplier
if pixel_values is None:
left_pad_tokens_per_prompt = calculate_pad_tokens_in_prompt(input_ids, logits_to_keep, trainer.processing_class.pad_token_id)
# Determine max_left_pad from precomputed logprobs shape for consistency
if old_logps is not None:
max_left_pad = old_logps.shape[1] - logits_to_keep
elif ref_logps is not None:
max_left_pad = ref_logps.shape[1] - logits_to_keep
else:
max_left_pad = torch.max(left_pad_tokens_per_prompt).item()
input_ids = left_pack_padding(input_ids, trainer.processing_class.pad_token_id)
completion_input_ids = input_ids[:, -(logits_to_keep +max_left_pad):]
completion_mask = create_completion_attention_mask(completion_input_ids, left_pad_tokens_per_prompt, max_left_pad, trainer.processing_class.pad_token_id).to(attention_mask.dtype)
if trainer.use_vllm and sampling_per_token_logps is not None and getattr(trainer, "vllm_importance_sampling_correction", False):
sampling_per_token_logps = align_logprobs_with_mask(sampling_per_token_logps, completion_mask)
else:
sampling_per_token_logps = None
completion_mask = align_completion_tool_mask(tool_mask, completion_mask)
attention_mask = input_ids != trainer.processing_class.pad_token_id
attention_mask = attention_mask.to(attention_mask.dtype)
else:
completion_input_ids = input_ids[:, -logits_to_keep:]
completion_mask = align_completion_tool_mask(tool_mask, completion_mask)
unwrapped_model = trainer.accelerator.unwrap_model(trainer.model, keep_fp32_wrapper = False)
for module in unwrapped_model.modules():
if hasattr(module, "_hf_hook") and hasattr(module._hf_hook, "io_same_decice"):
module._hf_hook.io_same_decice = False
pass
all_logprobs_list = []
def slice_sample_axis(value, start, end):
if value is None:
return None
return value[start:end]
import math
total_samples = input_ids.shape[0]
batch_size = math.ceil(total_samples / B)
if isinstance(num_images, torch.Tensor):
num_images = num_images.detach().cpu().reshape(-1).tolist()
if image_grid_thw is not None and pixel_values is not None and num_images is not None:
rows_per_image = image_grid_thw.prod(dim=-1)
rows_per_sample = torch.split(rows_per_image, num_images)
rows_per_sample = torch.stack([s.sum() for s in rows_per_sample])
cum_rows = torch.cat(
[
torch.tensor([0], device=rows_per_sample.device),
rows_per_sample.cumsum(0),
]
)
cum_imgs = torch.tensor([0] + num_images).cumsum(0)
else:
cum_rows = None
cum_imgs = None
input_ids_chunks = []
attention_mask_chunks = []
completion_ids_chunks = []
pixel_values_chunks = []
image_grid_thw_chunks = []
pixel_attention_mask_chunks = []
image_sizes_chunks = []
token_type_ids_chunks = []
mm_token_type_ids_chunks = []
current_pixel_idx = 0
#TRL 0.23.0 batching logic
for start in range(0, total_samples, batch_size):
end = min(start + batch_size, total_samples)
input_ids_chunks.append(input_ids[start:end])
attention_mask_chunks.append(attention_mask[start:end])
completion_ids_chunks.append(completion_input_ids[start:end])
image_sizes_chunks.append(slice_sample_axis(image_sizes, start, end))
token_type_ids_chunks.append(slice_sample_axis(token_type_ids, start, end))
mm_token_type_ids_chunks.append(
slice_sample_axis(mm_token_type_ids, start, end)
)
if image_grid_thw is not None and pixel_values is not None:
if num_images is None:
grid_slice = image_grid_thw[start:end]
batch_pixel_count = grid_slice.prod(dim=-1).sum().item()
start_pixel_idx = current_pixel_idx
end_pixel_idx = current_pixel_idx + batch_pixel_count
current_pixel_idx = end_pixel_idx
else:
start_pixel_idx = cum_rows[start].item()
end_pixel_idx = cum_rows[end].item()
img_start, img_end = cum_imgs[start], cum_imgs[end]
grid_slice = image_grid_thw[img_start:img_end]
image_grid_thw_chunks.append(grid_slice)
pixel_values_chunks.append(pixel_values[start_pixel_idx:end_pixel_idx])
if pixel_attention_mask is not None:
if pixel_attention_mask.shape[0] == pixel_values.shape[0]:
pixel_attention_mask_chunks.append(pixel_attention_mask[start_pixel_idx:end_pixel_idx])
else:
pixel_attention_mask_chunks.append(pixel_attention_mask[start:end])
else:
pixel_attention_mask_chunks.append(None)
else:
pixel_values_chunks.append(None)
image_grid_thw_chunks.append(None)
pixel_attention_mask_chunks.append(None)
zipped_inputs = zip(
input_ids_chunks,
attention_mask_chunks,
pixel_values_chunks,
image_grid_thw_chunks,
pixel_attention_mask_chunks,
image_sizes_chunks,
token_type_ids_chunks,
mm_token_type_ids_chunks,
completion_ids_chunks
)
if trainer._autocast_dtype is None:
autocaster = nullcontext()
else:
autocaster = torch.amp.autocast(device_type = trainer.model.device.type, dtype = trainer._autocast_dtype)
def to_device(tensor, device, non_blocking=True):
if tensor is None: return None
return tensor.to(device, non_blocking=non_blocking)
class Unsloth_Offloaded_Log_Softmax(torch.autograd.Function):
"""
Manual Gradient Checkpointing/CPU Offloading for Log Softmax.
"""
@staticmethod
def forward(ctx, hidden_states, lm_head, index, chunks,
logit_scale_multiply, logit_scale_divide,
logit_softcapping, temperature):
#Only the activations are needed so if we keep entire computational graph, keeps unnecessary memory on CPU so we detach it
ctx.saved_hidden_states = hidden_states.detach().contiguous().to("cpu", non_blocking=True)
ctx.device = hidden_states.device
ctx.dtype = hidden_states.dtype
ctx.lm_head = lm_head
ctx.lm_head_requires_grad = lm_head.requires_grad
ctx.index = index
ctx.args = (chunks, logit_scale_multiply, logit_scale_divide, logit_softcapping, temperature)
with torch.no_grad():
output = chunked_hidden_states_selective_log_softmax(
hidden_states, lm_head, index, *ctx.args
)
return output
@staticmethod
def backward(ctx, grad_output):
hidden_states = to_device(ctx.saved_hidden_states, ctx.device)
hidden_states = hidden_states.to(ctx.dtype)
hidden_states.requires_grad_(True)
lm_head = ctx.lm_head
# #Possibly redundant lines
# if ctx.lm_head_requires_grad:
# hidden_states.requires_grad_(True)
# else:
# lm_head = lm_head.detach()
index = ctx.index
with torch.enable_grad():
output = chunked_hidden_states_selective_log_softmax(
hidden_states, lm_head, index, *ctx.args
)
torch.autograd.backward(output, grad_output)
return (
hidden_states.grad,
lm_head.grad if ctx.lm_head_requires_grad else None,
None,
None,
None,
None,
None,
None,
)
def efficient_log_softmax(hidden_states, lm_head, index, chunks=32,
logit_scale_multiply=0.0, logit_scale_divide=0.0,
logit_softcapping=0.0, temperature=1, batch_size=8):
if (index.shape[1] <= 1024 and batch_size <= 8) or batch_size==1:
#We save a gigabyte or speed with the normal path under these specific conditions
return chunked_hidden_states_selective_log_softmax(
hidden_states,
lm_head,
index,
chunks,
logit_scale_multiply,
logit_scale_divide,
logit_softcapping,
temperature
)
else:
return Unsloth_Offloaded_Log_Softmax.apply(
hidden_states, lm_head, index, chunks,
logit_scale_multiply, logit_scale_divide,
logit_softcapping, temperature
)
def compute_logprobs_chunk(new_hidden_states_chunk, completion_ids, input_ids_chunk):
# Hidden states -> lm_head matmul path; raw logits -> skip matmul and
# skip scale/softcap (model forward already applied them).
chunks = input_ids_chunk.shape[0] * multiplier
if new_hidden_states_chunk.shape[-1] == lm_head.shape[1]:
return efficient_log_softmax(
new_hidden_states_chunk,
lm_head,
completion_ids,
chunks = chunks,
logit_scale_multiply = logit_scale_multiply,
logit_scale_divide = logit_scale_divide,
logit_softcapping = logit_softcapping,
temperature = temperature,
batch_size = B,
)
return chunked_selective_log_softmax(
new_hidden_states_chunk,
completion_ids,
temperature = temperature,
chunks = chunks,
)
for (
input_ids_chunk,
attention_mask_chunk,
pixel_values_chunk,
image_grid_thw_chunk,
pixel_attention_mask_chunk,
image_sizes_chunk,
token_type_ids_chunk,
mm_token_type_ids_chunk,
completion_ids
) in zipped_inputs:
_extra_vision_kwargs = {}
if token_type_ids_chunk is not None:
_extra_vision_kwargs["token_type_ids"] = token_type_ids_chunk
if mm_token_type_ids_chunk is not None:
_extra_vision_kwargs["mm_token_type_ids"] = mm_token_type_ids_chunk
with autocaster:
if pixel_values is None:
new_hidden_states_chunk = unwrapped_model(
input_ids = input_ids_chunk,
attention_mask = attention_mask_chunk,
pixel_values = pixel_values_chunk,
image_grid_thw = image_grid_thw_chunk,
pixel_attention_mask = pixel_attention_mask_chunk,
image_sizes = image_sizes_chunk,
**_extra_vision_kwargs,
).logits
new_hidden_states_chunk = new_hidden_states_chunk[:, -(logits_to_keep + max_left_pad + 1): , :]
new_hidden_states_chunk = new_hidden_states_chunk[:, :-1, :]
logprobs_chunk = compute_logprobs_chunk(new_hidden_states_chunk, completion_ids, input_ids_chunk)
else:
new_hidden_states_chunk = unwrapped_model(
input_ids = input_ids_chunk,
attention_mask = attention_mask_chunk,
pixel_values = pixel_values_chunk,
image_grid_thw = image_grid_thw_chunk,
pixel_attention_mask = pixel_attention_mask_chunk,
image_sizes = image_sizes_chunk,
logits_to_keep = logits_to_keep + 1,
**_extra_vision_kwargs,
).logits
new_hidden_states_chunk = new_hidden_states_chunk[:, :-1, :]
logprobs_chunk = compute_logprobs_chunk(new_hidden_states_chunk, completion_ids, input_ids_chunk)
#This is needed to avoid race conditions with GPT OSS offload_embbed=True
#However, it seems that this line does not slow down or disrupt models.
device_synchronize()
all_logprobs_list.append(logprobs_chunk)
new_logprobs = torch.cat(all_logprobs_list, dim=0)
with autocaster:
loss, completion_length, mean_kl, delta, flat_is_ratio, coef_1 = UnslothEfficientGRPO.apply(
new_logprobs,
old_logps,
ref_logps,
sampling_per_token_logps,
lm_head,
completion_input_ids,
completion_mask,
advantages,
trainer.beta,
trainer.accelerator.scaler,
1,
kwargs
)
# Must force not returning hidden states but logits otherwise gibberish
os.environ["UNSLOTH_RETURN_HIDDEN_STATES"] = "0"
return loss, completion_length, mean_kl, delta, flat_is_ratio, coef_1, completion_mask
# Old non efficient code path
new_logits = torch.matmul(new_hidden_states, lm_head.t())
new_logits = new_logits[:, :-1, :] # exclude the last logit: it corresponds to the next token pred
old_logits = torch.matmul(old_hidden_states, lm_head.t())
old_logits = old_logits[:, :-1, :] # exclude the last logit: it corresponds to the next token pred
loss, completion_length, mean_kl = grpo_compute_loss(
old_logits,
new_logits,
completion_input_ids,
completion_mask,
trainer.beta,
advantages,
)
return loss, completion_length, mean_kl
pass
@torch.compile(dynamic = True, fullgraph = True, options = torch_compile_options)
def grpo_compute_loss_slow(
ref,
new,
old,
sampling_per_token_logps,
input_ids,
mask,
beta,
advantages,
**kwargs
):
# All Unsloth Zoo code licensed under AGPL3
# Set defaults for optional arguments
loss_type = kwargs.get("loss_type", "grpo")
epsilon_low = kwargs.get("epsilon_low", 0.2)
epsilon_high = kwargs.get("epsilon_high", 0.2)
max_completion_length = kwargs.get("max_completion_length", 8192)
delta = kwargs.get("delta", None)
importance_sampling_level = kwargs.get("importance_sampling_level", "token")
num_items_in_batch = kwargs.get("num_items_in_batch", None)
current_gradient_accumulation_steps = kwargs.get("current_gradient_accumulation_steps", 1)
num_processes = kwargs.get("num_processes", 1)
use_vllm = kwargs.get("use_vllm", False)
vllm_importance_sampling_cap = kwargs.get("vllm_importance_sampling_cap", 2.0)
get_sapo_token_loss = kwargs.get("get_sapo_token_loss", None)
sapo_temperature_pos = kwargs.get("sapo_temperature_pos", 1.0)
sapo_temperature_neg = kwargs.get("sapo_temperature_neg", 1.05)
get_gamma_weights = kwargs.get("get_gamma_weights", None)
vespo_k_pos = kwargs.get("vespo_k_pos", 2.0)
vespo_lambda_pos = kwargs.get("vespo_lambda_pos", 3.0)
vespo_k_neg = kwargs.get("vespo_k_neg", 3.0)
vespo_lambda_neg = kwargs.get("vespo_lambda_neg", 2.0)
get_off_policy_mask = kwargs.get("get_off_policy_mask", None)
off_policy_mask_threshold = kwargs.get("off_policy_mask_threshold", None)
input_ids = input_ids.unsqueeze(-1)
if advantages.dim() == 1:
advantages = advantages.unsqueeze(1)
if off_policy_mask_threshold is not None:
off_policy_mask = get_off_policy_mask(
advantages=advantages,
per_token_logps=new,
old_per_token_logps=old,
mask=mask,
off_policy_threshold=off_policy_mask_threshold,
)
with torch.no_grad():
if use_vllm and sampling_per_token_logps is not None:
#must filter out extra prompt tokens in begining after making input_ids left padded
importance_sampling_ratio = torch.exp((old * mask) - sampling_per_token_logps)
importance_sampling_ratio = torch.clamp(
importance_sampling_ratio, max=vllm_importance_sampling_cap
)
pass
# Must detach - otherwise gradients are not propagated correctly!
# exp(x - x) == 1
# loss_i = torch.exp(new - new.detach()) * advantages.unsqueeze(1)
if old is not None:
log_ratio = new - old
else:
log_ratio = new - new.detach()
if importance_sampling_level == "token":
log_importance_weights = log_ratio
elif importance_sampling_level == "sequence":
log_importance_weights = (log_ratio * mask).sum(-1) / mask.sum(-1).clamp(min=1.0)
log_importance_weights = log_importance_weights.unsqueeze(-1)
else:
raise ValueError(
f"Unknown importance sampling level: {importance_sampling_level}. Possible values are 'token' "
"and 'sequence'."
)
coef_1 = torch.exp(log_importance_weights)
# Reverse KL
# Note that this is a low variance low bias estimator for the KL divergence as used in GRPO paper
if beta != 0.0:
kl_i = torch.exp(ref - new) - (ref - new) - 1.0
else:
# set kl_i to a tensor of zeros with the correct shape
if importance_sampling_level == "sequence":
kl_i = new.new_zeros(new.size(0), 1)
else:
kl_i = torch.zeros_like(new)
# Full correct reverse KL divergence?? Missing term maybe?
# kl_i = torch.exp(new) * kl_i
# Below is forward KL (normal KL)
# kl_i = torch.exp(old) * (old - new)
if loss_type == "cispo":
clamped_ratios = torch.clamp(coef_1, max=epsilon_high).detach()
loss_i = -clamped_ratios * advantages * new
#breakpoint()
elif loss_type in ["grpo", "bnpo", "dr_grpo", "dapo"]:
coef_2 = torch.clamp(coef_1, 1 - epsilon_low, 1 + epsilon_high)
if delta is not None:
loss_1 = torch.clamp(coef_1, max=delta) * advantages
else:
loss_1 = coef_1 * advantages
pass
loss_2 = coef_2 * advantages
loss_i = -torch.min(loss_1, loss_2)
elif loss_type == "sapo":
if get_sapo_token_loss is None:
raise Exception(f"sapo is only available in TRL 0.26.0+")
loss_i = torch.empty_like(coef_1)
positive_advantages_mask = advantages.repeat([1, coef_1.shape[1]]) > 0
#since we have n_chunks some tensors may error if they dont have elements in them
if coef_1[positive_advantages_mask].numel() != 0:
loss_i[positive_advantages_mask] = get_sapo_token_loss(
coef_1[positive_advantages_mask], sapo_temperature_pos
)
if coef_1[~positive_advantages_mask].numel() != 0:
loss_i[~positive_advantages_mask] = get_sapo_token_loss(
coef_1[~positive_advantages_mask], sapo_temperature_neg
)
loss_i = -loss_i * advantages
elif loss_type == "vespo":
if get_gamma_weights is None:
raise Exception("vespo is only available in TRL 0.26.0+")
phi_seq = get_gamma_weights(
advantages=advantages,
log_ratio_per_token=log_ratio,
mask=mask,
importance_sampling_ratio=kwargs.get("importance_sampling_ratio"),
k_pos=vespo_k_pos,
lambda_pos=vespo_lambda_pos,
k_neg=vespo_k_neg,
lambda_neg=vespo_lambda_neg,
)
loss_i = -phi_seq * advantages * new
else:
raise ValueError(f"Unknown loss type: {loss_type}")
if off_policy_mask_threshold is not None:
loss_i = loss_i * off_policy_mask
if use_vllm and sampling_per_token_logps is not None:
loss_i = loss_i * importance_sampling_ratio
#delta for metric
with torch.no_grad():
delta = torch.abs(old - sampling_per_token_logps)
delta = delta * mask
flat_is_ratio = importance_sampling_ratio * mask
else:
delta = torch.tensor([]).detach()
flat_is_ratio = torch.tensor([]).detach()
if beta != 0.0:
loss_i = loss_i + beta * kl_i
mask = mask.to(torch.float32)
n_mask_per_reward = mask.sum(1)
# https://github.com/huggingface/trl/blob/e8b8499f1f8d76838155b515e414ee98f757d6d5/trl/trainer/grpo_trainer.py#L1624
if loss_type in ["grpo", "sapo"]:
loss = ((loss_i * mask).sum(-1) / mask.sum(-1).clamp(min=1.0)).mean()
loss = loss / current_gradient_accumulation_steps
elif loss_type == "bnpo":
loss = (loss_i * mask).sum() / mask.sum().clamp(min=1.0)
loss = loss / current_gradient_accumulation_steps
elif loss_type == "dr_grpo":
loss = (loss_i * mask).sum() / (loss_i.size(0) * max_completion_length)
loss = loss / current_gradient_accumulation_steps
elif loss_type in ["cispo", "dapo", "vespo"]:
normalizer = num_items_in_batch/ num_processes
loss = (loss_i * mask).sum() / normalizer
else:
raise ValueError(f"Unknown loss type: {loss_type}")
# loss = (loss_i * mask).sum() / mask.sum()
# Get metrics as well which are folded
def masked_batch_mean(x):
with torch.inference_mode():
completion_length = n_mask_per_reward.mean()
if x.shape[1] == 1: # when importance_sampling_level == "sequence"
return completion_length, x.mean()
else:
mean_kl_per_reward = (x * mask).sum(1) / n_mask_per_reward
mean_kl = mean_kl_per_reward.mean()
return completion_length, mean_kl
completion_length, mean_kl = masked_batch_mean(kl_i)
return loss, completion_length, mean_kl, delta, flat_is_ratio, coef_1, mask
def grpo_update_SamplingParams(SamplingParams, generation_kwargs, vllm_sampling_params = None):
good_sampling_params_keys = inspect.signature(SamplingParams).parameters.keys()
# Filter generation_kwargs
new_generation_kwargs = {}
for key in generation_kwargs.keys():
if key in good_sampling_params_keys:
new_generation_kwargs[key] = generation_kwargs[key]
generation_kwargs = new_generation_kwargs
if vllm_sampling_params is not None:
for key in good_sampling_params_keys:
if hasattr(vllm_sampling_params, key):
overwrited_key = getattr(vllm_sampling_params, key)
if overwrited_key is not None and (type(overwrited_key) in (list, tuple,) and len(overwrited_key) != 0):
generation_kwargs[key] = overwrited_key
return generation_kwargs
def _get_inference_mode_context_manager(model: torch.nn.Module):
"""
If the state dict was quantized using torchao, we will run into
the following error when calling ops like aten.t() in inference mode.
This is a bug in PyTorch that affects all tensor subclasses.
Cannot set version_counter for inference tensor
For now, we work around this issue by using `torch.no_grad()` in this case.
See https://github.com/pytorch/pytorch/issues/164872 for more details.
Otherwise, just return `torch.inference_mode()`.
"""
torchao_config = getattr(model, "torchao_config", None)
if torchao_config is not None and torchao_config.qat_scheme is None:
return torch.no_grad()
else:
return torch.inference_mode()
def vLLMSamplingParams(**kwargs):
from vllm import SamplingParams
sampling_params = SamplingParams(**kwargs)
sampling_params._set_kwargs = kwargs
return sampling_params
@dataclass
class UnslothGRPOConfig(GRPOConfig):
"""
Configuration class for the [`GRPOTrainer`].
This class includes only the parameters that are specific to GRPO training. For a full list of training arguments,
please refer to the [`~transformers.TrainingArguments`] documentation. Note that default values in this class may
differ from those in [`~transformers.TrainingArguments`].
Using [`~transformers.HfArgumentParser`] we can turn this class into
[argparse](https://docs.python.org/3/library/argparse#module-argparse) arguments that can be specified on the
command line.
Parameters:
> Parameters that control the model and reference model
model_init_kwargs (`str`, `dict[str, Any]`, *optional*):
Keyword arguments for [`~transformers.AutoModelForCausalLM.from_pretrained`], used when the `model`
argument of the [`GRPOTrainer`] is provided as a string.
disable_dropout (`bool`, *optional*, defaults to `False`):
Whether to disable dropout in the model. This is useful for training with a reference model, as it prevents
the model from generating different logprobs for the same input.
> Parameters that control the data preprocessing
remove_unused_columns (`bool`, *optional*, defaults to `False`):
Whether to only keep the column `"prompt"` in the dataset. If you use a custom reward function that
requires any column other than `"prompts"` and `"completions"`, you should keep this to `False`.
max_prompt_length (`int` or `None`, *optional*, defaults to `512`):
Maximum length of the prompt. If the prompt is longer than this value, it will be truncated left.
num_generations (`int` or `None`, *optional*, defaults to `8`):
Number of generations per prompt to sample. The effective batch size (num_processes * per_device_batch_size
* gradient_accumulation_steps) must be evenly divisible by this value.
max_completion_length (`int` or `None`, *optional*, defaults to `256`):
Maximum length of the generated completion.
ds3_gather_for_generation (`bool`, *optional*, defaults to `True`):
This setting applies to DeepSpeed ZeRO-3. If enabled, the policy model weights are gathered for generation,
improving generation speed. However, disabling this option allows training models that exceed the VRAM
capacity of a single GPU, albeit at the cost of slower generation. Disabling this option is not compatible
with vLLM generation.
shuffle_dataset (`bool`, *optional*, defaults to `True`):
Whether to shuffle the training dataset.
> Parameters that control generation
generation_batch_size: (`int`, *optional*):
Batch size to use for generation. If `None`, it defaults to the effective training batch size:
`per_device_train_batch_size * num_processes * steps_per_generation`. In other words, there is one
generation batch processed per optimization step. Mutually exclusive with `steps_per_generation`.
steps_per_generation: (`int`, *optional*):
Number of steps per generation. If `None`, it defaults to `gradient_accumulation_steps`. Mutually exclusive
with `generation_batch_size`.
temperature (`float`, defaults to `1.0`):
Temperature for sampling. The higher the temperature, the more random the completions.
top_p (`float`, *optional*, defaults to `1.0`):
Float that controls the cumulative probability of the top tokens to consider. Must be in (0, 1]. Set to
`1.0` to consider all tokens.
top_k (`int`, *optional*):
Number of highest probability vocabulary tokens to keep for top-k-filtering. If `None`, top-k-filtering is
disabled and all tokens are considered.
min_p (`float`, *optional*):
Minimum token probability, which will be scaled by the probability of the most likely token. It must be a
value between `0.0` and `1.0`. Typical values are in the `0.01-0.2` range.
repetition_penalty (`float`, *optional*, defaults to `1.0`):
Float that penalizes new tokens based on whether they appear in the prompt and the generated text so far.
Values > `1.0` encourage the model to use new tokens, while values < `1.0` encourage the model to repeat
tokens.
use_transformers_paged (`bool`, *optional*, defaults to `False`):
Whether to use the `transformers` paged implementation for generation. If set to `True`, the `transformers`
paged implementation will be used for generation instead of the default padded implementation. This
parameter is only effective when `use_vllm` is set to `False`.
cache_implementation (`str`, *optional*):
Implementation of the cache method for faster generation when `use_vllm` is set to `False`.
generation_kwargs (`dict[str, Any]`, *optional*):
Additional keyword arguments to pass to [`~transformers.GenerationConfig`] (if using transformers) or
`SamplingParams` (if using vLLM) when sampling completions. This can be used to further customize the
generation behavior, such as setting `suppress_tokens`, `num_beams`, etc. If it contains keys that conflict
with the other generation parameters (like `min_p`, `top_p`, etc.), they will override them.
> Parameters that control generation acceleration powered by vLLM
use_vllm (`bool`, *optional*, defaults to `False`):
Whether to use vLLM for generating completions. If set to `True`, the trainer will use vLLM for generation
instead of the default model.generate(). Requires `vllm` to be installed.
vllm_mode (`str`, *optional*, defaults to `"server"`):
Mode to use for vLLM integration when `use_vllm` is set to `True`. Must be one of `"server"` or
`"colocate"`.
- `"server"`: The trainer will send generation requests to a separate vLLM server. Make sure a TRL vLLM
server is running (start with `trl vllm-serve`).
- `"colocate"`: vLLM will run in the same process and share the training GPUs. This avoids the need for a
separate server but may cause resource contention with training.
vllm_model_impl (`str`, *optional*, defaults to `"vllm"`):
Model implementation to use for vLLM. Must be one of `"transformers"` or `"vllm"`. `"transformers"`: Use
the `transformers` backend for model implementation. `"vllm"`: Use the `vllm` library for model
implementation.
vllm_guided_decoding_regex (`str`, *optional*):
Regex for vLLM guided decoding. If `None` (default), guided decoding is disabled.
> Parameters that control the vLLM server (only used when `vllm_mode` is `"server"`)
vllm_server_base_url (`str`, *optional*):
Base URL for the vLLM server (e.g., `"http://localhost:8000"`). If provided, `vllm_server_host` and
`vllm_server_port` are ignored.
vllm_server_host (`str`, *optional*, defaults to `"0.0.0.0"`):
Host of the vLLM server to connect to. Ignored if `vllm_server_base_url` is provided.
vllm_server_port (`int`, *optional*, defaults to `8000`):
Port of the vLLM server to connect to. Ignored if `vllm_server_base_url` is provided.
vllm_server_timeout (`float`, *optional*, defaults to `240.0`):
Total timeout duration in seconds to wait for the vLLM server to be up. If the server is not up after the
timeout, a `ConnectionError` is raised.
> Parameters that control colocated vLLM execution (only used when `vllm_mode` is `"colocate"`)
vllm_gpu_memory_utilization (`float`, *optional*, defaults to `0.3`):
Control the GPU memory utilization for vLLM. This setting only applies when `vllm_mode` is set to
`"colocate"`. If you are using `vllm_mode="server"`, this parameter must be passed separately when
launching the vLLM server via the `--vllm_gpu_memory_utilization` flag.
vllm_tensor_parallel_size (`int`, *optional*, defaults to `1`):
Control the tensor parallel size for vLLM. This setting only applies when `vllm_mode` is set to
`"colocate"`. If you are using `vllm_mode="server"`, this parameter must be passed separately when
launching the vLLM server via the `--vllm_tensor_parallel_size` flag.
vllm_enable_sleep_mode (`bool`, *optional*, defaults to `False`):
Whether to enable sleep mode for vLLM. If `True`, vLLM will sleep during the optimization step and woken
for weight sync and generation.
> Parameters that control the training
beta (`float`, *optional*, defaults to `0.0`):
KL coefficient. If `0.0` (default), the reference model is not loaded, reducing memory usage and improving
training speed.
num_iterations (`int`, *optional*, defaults to `1`):
Number of iterations per batch (denoted as μ in the algorithm).
epsilon (`float`, *optional*, defaults to `0.2`):
Epsilon value for clipping.
delta (`float`, *optional*):
Enables the upper clipping bound in two-sided GRPO loss when set to a float. If `None` (default), standard
GRPO clipping is used. Recommended to be greater than `1 + ε` when enabled. This method is introduced in
the [INTELLECT-2 tech report](https://huggingface.co/papers/2505.07291).
epsilon_high (`float`, *optional*):
Upper-bound epsilon value for clipping. If not specified, it defaults to the same value as the lower-bound
specified in argument `epsilon`. Paper [DAPO](https://huggingface.co/papers/2503.14476) recommends `0.28`.
importance_sampling_level (`str`, *optional*, defaults to `"token"`):
Controls whether importance sampling ratios are computed at the `"token"` or `"sequence"` level. `"token"`
keeps the raw per-token log-probability ratios (one weight per token). `"sequence"` averages the
log-probability ratios across valid tokens to produce a single ratio per sequence. The [GSPO
paper](https://huggingface.co/papers/2507.18071) shows that sequence-level sampling often yields more
stable training and better alignment with sequence-level rewards.
reward_weights (`list[float]`, *optional*):
Weights for each reward function. Must match the number of reward functions. If `None`, all rewards are
weighted equally with weight `1.0`.
scale_rewards (`str` or `bool`, *optional*, defaults to `"group"`):
Specifies the scaling strategy for rewards. Supported values are:
- `True` or `"group"` (default): rewards are scaled by the standard deviation within each group, ensuring
unit variance within a group.
- `"batch"`: rewards are scaled by the standard deviation across the entire batch, as recommended in the
[PPO Lite paper](https://huggingface.co/papers/2508.08221).
- `False` or `"none"`: no scaling is applied. The [Dr. GRPO
paper](https://huggingface.co/papers/2503.20783) recommends not scaling rewards, as scaling by the
standard deviation introduces a question-level difficulty bias.
loss_type (`str`, *optional*, defaults to `"dapo"`):
Specifies the loss formulation to use. Supported values are:
- `"grpo"`: Aggregates token-level losses by normalizing over sequence length. Not recommended due to
length bias—this approach tends to prefer shorter completions with positive advantages and longer ones
with negative advantages.
- `"dr_grpo"`: Aggregates token-level losses by normalizing with a global constant. This method was
introduced in the [Dr. GRPO paper](https://huggingface.co/papers/2503.20783) to eliminate length bias.
The value of the constant corresponds to `max_completion_length`.
- `"dapo"` (default): Aggregates token-level losses by normalizing with the number of active token in the
global accumulated batch. This method was introduced in the [DAPO
paper](https://huggingface.co/papers/2503.14476) to eliminate length bias.
- `"bnpo"`: Aggregates token-level losses by normalizing with the number of active token in the local
batch. Note that normalization is performed over the local batch only, so results may slightly vary
depending on the local batch size, despite a constant effective batch size. When using
`per_device_train_batch_size==1`, the loss is equivalent to the GRPO loss.
mask_truncated_completions (`bool`, *optional*, defaults to `False`):
When enabled, truncated completions are excluded from the loss calculation, preventing them from being
incorrectly penalized and introducing noise during training. According to the
[DAPO](https://huggingface.co/papers/2503.14476) paper, this is a good practice for training stability.
sync_ref_model (`bool`, *optional*, defaults to `False`):
Whether to synchronize the reference model with the active model every `ref_model_sync_steps` steps, using
the `ref_model_mixup_alpha` parameter. This synchronization originates from the
[TR-DPO](https://huggingface.co/papers/2404.09656) paper.
ref_model_mixup_alpha (`float`, *optional*, defaults to `0.6`):
α parameter from the [TR-DPO](https://huggingface.co/papers/2404.09656) paper, which controls the mix
between the current policy and the previous reference policy during updates. The reference policy is
updated according to the equation: `π_ref = α * π_θ + (1 - α) * π_ref_prev`. To use this parameter, you
must set `sync_ref_model=True`.
ref_model_sync_steps (`int`, *optional*, defaults to `512`):
τ parameter from the [TR-DPO](https://huggingface.co/papers/2404.09656) paper, which determines how
frequently the current policy is synchronized with the reference policy. To use this parameter, you must
set `sync_ref_model=True`.
top_entropy_quantile (`float`, *optional*, defaults to `1.0`):
ρ parameter from [Beyond the 80/20 Rule](https://huggingface.co/papers/2506.01939). Keeps in the policy
loss term only the top-ρ quantile of tokens by entropy of the probability distribution at each sequence
position, improving results. Range: `[0.0-1.0]`. A value of `0.0` masks all but the highest entropy token;
`1.0` keeps all tokens. The paper recommends a value of `0.2`. If used with
`mask_truncated_completions=True`, only tokens from non-truncated completions are considered.
use_liger_loss (`bool`, *optional*, defaults to `False`):
Whether to use the Liger GRPO loss.
vllm_importance_sampling_correction (`bool`, *optional*, defaults to `True`):
Whether to apply Truncated Importance Sampling (TIS) between vLLM completion logprobs and recomputed
logprobs. [Your Efficient RL Framework Secretly Brings You Off-Policy RL
Training](https://fengyao.notion.site/off-policy-rl) highlights that using a separate generation framework
(such as vLLM) can introduce off-policy effects due to subtle implementation differences between generation
and training backends. TIS is proposed as a remedy for this issue.
vllm_importance_sampling_cap (`float`, *optional*, defaults to `2.0`):
Truncation parameter C for Truncated Importance Sampling (TIS). This sets an upper bound on the importance
sampling ratio, improving training stability.
> Parameters that control the logging
log_completions (`bool`, *optional*, defaults to `False`):
Whether to log a sample of (prompt, completion) pairs every `logging_steps` steps. If `rich` is installed,
it prints the sample. If `wandb` logging is enabled, it logs it to `wandb`.
num_completions_to_print (`int`, *optional*):
Number of completions to print with `rich`. If `None`, all completions are logged.
wandb_log_unique_prompts (`bool`, *optional*, defaults to `False`):
Whether to log unique prompts in wandb. If `True`, only unique prompts are logged. If `False`, all prompts
are logged.
"""
vllm_sampling_params: Optional[Any] = field(
default = None,
metadata = {'help': 'vLLM SamplingParams'},
)
unsloth_num_chunks : Optional[int] = field(
default = -1,
metadata = {'help': 'Chunk size to reduce memory usage. -1 is most efficient.'},
)
unsloth_logit_chunk_multiplier : Optional[int] = field(
default = None,
metadata = {'help': 'Multiplier for chunked logit computations.'},
)
unsloth_grpo_mini_batch : Optional[int] = field(
default = None,
metadata = {'help': 'Mini batch size for GRPO hidden state accumulation. Default is None unless user defines it.'},
)
def __init__(
self,
output_dir = None,
per_device_train_batch_size = 4,
num_train_epochs = 3.0,
max_steps = -1,
learning_rate = 5e-05,
lr_scheduler_type = 'linear',
lr_scheduler_kwargs = None,
warmup_steps = 0.1,
optim = 'adamw_8bit',
optim_args = None,
weight_decay = 0.001,
adam_beta1 = 0.9,
adam_beta2 = 0.999,
adam_epsilon = 1e-08,
optim_target_modules = None,
gradient_accumulation_steps = 2,
average_tokens_across_devices = True,
max_grad_norm = 1.0,
label_smoothing_factor = 0.0,
bf16 = False,
fp16 = False,
bf16_full_eval = False,
fp16_full_eval = False,
tf32 = None,
gradient_checkpointing = True,
gradient_checkpointing_kwargs = None,
torch_compile = False,
torch_compile_backend = None,
torch_compile_mode = None,
use_liger_kernel = False,
liger_kernel_config = None,
use_cache = False,
neftune_noise_alpha = None,
torch_empty_cache_steps = 250,
auto_find_batch_size = False,
logging_strategy = 'steps',
logging_steps = 1,
logging_first_step = False,
log_on_each_node = True,
logging_nan_inf_filter = False,
include_num_input_tokens_seen = False,
log_level = 'passive',
log_level_replica = 'warning',
disable_tqdm = None,
report_to = 'none',
run_name = None,
project = 'huggingface',
trackio_space_id = 'trackio',
eval_strategy = 'no',
eval_steps = None,
eval_delay = 0,
per_device_eval_batch_size = 4,
prediction_loss_only = False,
eval_on_start = False,
eval_do_concat_batches = True,
eval_use_gather_object = False,
eval_accumulation_steps = 2,
batch_eval_metrics = False,
save_only_model = False,
save_strategy = 'steps',
save_steps = 500,
save_on_each_node = False,
save_total_limit = None,
enable_jit_checkpoint = False,
push_to_hub = False,
hub_token = None,
hub_private_repo = None,
hub_model_id = None,
hub_strategy = 'every_save',
hub_always_push = False,
hub_revision = None,
load_best_model_at_end = False,
metric_for_best_model = None,
greater_is_better = None,
ignore_data_skip = False,
restore_callback_states_from_checkpoint = False,
full_determinism = False,
seed = 3407,
data_seed = 3407,
use_cpu = False,
accelerator_config = None,
parallelism_config = None,
dataloader_drop_last = False,
dataloader_num_workers = 0,
dataloader_pin_memory = True,
dataloader_persistent_workers = False,
dataloader_prefetch_factor = None,
remove_unused_columns = False,
label_names = None,
train_sampling_strategy = 'random',
length_column_name = 'length',
ddp_find_unused_parameters = None,
ddp_bucket_cap_mb = None,
ddp_broadcast_buffers = None,
ddp_backend = None,
ddp_timeout = 1800,
fsdp = None,
fsdp_config = None,
deepspeed = None,
debug = '',
skip_memory_metrics = True,
do_train = False,
do_eval = False,
do_predict = False,
resume_from_checkpoint = None,
warmup_ratio = None,
logging_dir = None,
local_rank = -1,
model_init_kwargs = None,
disable_dropout = False,
max_prompt_length = 512,
num_generations = 8,
max_completion_length = 256,
ds3_gather_for_generation = True,
shuffle_dataset = True,
generation_batch_size = None,
steps_per_generation = None,
temperature = 1.0,
top_p = 1.0,
top_k = None,
min_p = None,
generation_kwargs = {},
repetition_penalty = 1.0,
use_transformers_paged = False,
cache_implementation = None,
use_vllm = False,
vllm_mode = 'colocate',
vllm_model_impl = 'vllm',
vllm_enable_sleep_mode = False,
vllm_guided_decoding_regex = None,
vllm_server_base_url = None,
vllm_server_host = '0.0.0.0',
vllm_server_port = 8000,
vllm_server_timeout = 240.0,
vllm_gpu_memory_utilization = 0.3,
vllm_tensor_parallel_size = 1,
beta = 0.001,
num_iterations = 1,
epsilon = 0.2,
delta = None,
epsilon_high = None,
importance_sampling_level = 'token',
reward_weights = None,
scale_rewards = 'group',
loss_type = 'bnpo',
mask_truncated_completions = False,
sync_ref_model = False,
ref_model_mixup_alpha = 0.6,
ref_model_sync_steps = 512,
top_entropy_quantile = 1.0,
use_liger_loss = False,
vllm_importance_sampling_correction = False,
vllm_importance_sampling_cap = 2.0,
log_completions = False,
num_completions_to_print = None,
wandb_log_unique_prompts = False,
vllm_sampling_params = None,
unsloth_num_chunks = -1,
unsloth_logit_chunk_multiplier = None,
unsloth_grpo_mini_batch = None,
**kwargs,
):
if learning_rate < 1e-7: print(f'Unsloth: Your learning rate of `{learning_rate}` is too small and less than 1e-7! Consider increasing it, otherwise gradient updates will be close to 0!')
if learning_rate > 1: print(f'Unsloth: Your learning rate of `{learning_rate}` is way too larger > 1! Consider decreasing it to 1e-1, otherwise gradient updates will explode!')
if num_train_epochs is None:
num_train_epochs = 3.0 # Default to 3 epochs if None, max_steps will override
if output_dir is None and save_strategy == 'steps' and save_steps == 500:
output_dir = 'unsloth_training_checkpoints'
save_strategy = 'no'
if loss_type.lower() == 'dr_grpo':
loss_type = 'dr_grpo'
elif loss_type.lower() == 'dapo':
loss_type = 'dapo'
if loss_type.lower() == 'dr_grpo':
if scale_rewards == None:
scale_rewards = True
elif scale_rewards == True:
print('Unsloth: The Dr GRPO paper recommends setting `scale_rewards` to False! Will override. Set it to `None` to force False.')
scale_rewards = False
elif loss_type.lower() == 'dapo':
if mask_truncated_completions != True:
print('Unsloth: The DAPO paper recommends `mask_truncated_completions = True` - we will set it.')
if epsilon_high != 0.28:
print('Unsloth: The DAPO paper recommends `epsilon_high = 0.28` - we will set it.')
if beta != 0.0:
print(f'[WARNING] Unsloth: The DAPO paper recommends setting `beta = 0.0` to remove the KL term - You have set it to {beta}.')
mask_truncated_completions = True
epsilon_high = 0.28
if steps_per_generation is None and generation_batch_size is None:
ga = gradient_accumulation_steps
world_size = int(os.environ.get('WORLD_SIZE', '1'))
if (ga * world_size * per_device_train_batch_size) % num_generations != 0:
print('Unsloth: We now expect `per_device_train_batch_size` * `gradient_accumulation_steps` * `world_size` to be a multiple of `num_generations`.\nWe will change the batch size of ' + str(per_device_train_batch_size) + ' to the `num_generations` of ' + str(num_generations))
per_device_train_batch_size = num_generations
if temperature <= 0:
raise ValueError('Unsloth: Please set a positive non-zero temperature since your results will be wrong.')
elif temperature >= 10:
raise ValueError('Unsloth: Please set a positive non-zero temperature less than 10, since sampling will be quite erratic.')
if use_vllm and (top_k is None or top_k == 0): top_k = -1
super().__init__(
output_dir = output_dir,
per_device_train_batch_size = per_device_train_batch_size,
num_train_epochs = num_train_epochs,
max_steps = max_steps,
learning_rate = learning_rate,
lr_scheduler_type = lr_scheduler_type,
lr_scheduler_kwargs = lr_scheduler_kwargs,
warmup_steps = warmup_steps,
optim = optim,
optim_args = optim_args,
weight_decay = weight_decay,
adam_beta1 = adam_beta1,
adam_beta2 = adam_beta2,
adam_epsilon = adam_epsilon,
optim_target_modules = optim_target_modules,
gradient_accumulation_steps = gradient_accumulation_steps,
average_tokens_across_devices = average_tokens_across_devices,
max_grad_norm = max_grad_norm,
label_smoothing_factor = label_smoothing_factor,
bf16 = bf16,
fp16 = fp16,
bf16_full_eval = bf16_full_eval,
fp16_full_eval = fp16_full_eval,
tf32 = tf32,
gradient_checkpointing = gradient_checkpointing,
gradient_checkpointing_kwargs = gradient_checkpointing_kwargs,
torch_compile = torch_compile,
torch_compile_backend = torch_compile_backend,
torch_compile_mode = torch_compile_mode,
use_liger_kernel = use_liger_kernel,
liger_kernel_config = liger_kernel_config,
use_cache = use_cache,
neftune_noise_alpha = neftune_noise_alpha,
torch_empty_cache_steps = torch_empty_cache_steps,
auto_find_batch_size = auto_find_batch_size,
logging_strategy = logging_strategy,
logging_steps = logging_steps,
logging_first_step = logging_first_step,
log_on_each_node = log_on_each_node,
logging_nan_inf_filter = logging_nan_inf_filter,
include_num_input_tokens_seen = include_num_input_tokens_seen,
log_level = log_level,
log_level_replica = log_level_replica,
disable_tqdm = disable_tqdm,
report_to = report_to,
run_name = run_name,
project = project,
trackio_space_id = trackio_space_id,
eval_strategy = eval_strategy,
eval_steps = eval_steps,
eval_delay = eval_delay,
per_device_eval_batch_size = per_device_eval_batch_size,
prediction_loss_only = prediction_loss_only,
eval_on_start = eval_on_start,
eval_do_concat_batches = eval_do_concat_batches,
eval_use_gather_object = eval_use_gather_object,
eval_accumulation_steps = eval_accumulation_steps,
batch_eval_metrics = batch_eval_metrics,
save_only_model = save_only_model,
save_strategy = save_strategy,
save_steps = save_steps,
save_on_each_node = save_on_each_node,
save_total_limit = save_total_limit,
enable_jit_checkpoint = enable_jit_checkpoint,
push_to_hub = push_to_hub,
hub_token = hub_token,
hub_private_repo = hub_private_repo,
hub_model_id = hub_model_id,
hub_strategy = hub_strategy,
hub_always_push = hub_always_push,
hub_revision = hub_revision,
load_best_model_at_end = load_best_model_at_end,
metric_for_best_model = metric_for_best_model,
greater_is_better = greater_is_better,
ignore_data_skip = ignore_data_skip,
restore_callback_states_from_checkpoint = restore_callback_states_from_checkpoint,
full_determinism = full_determinism,
seed = seed,
data_seed = data_seed,
use_cpu = use_cpu,
accelerator_config = accelerator_config,
parallelism_config = parallelism_config,
dataloader_drop_last = dataloader_drop_last,
dataloader_num_workers = dataloader_num_workers,
dataloader_pin_memory = dataloader_pin_memory,
dataloader_persistent_workers = dataloader_persistent_workers,
dataloader_prefetch_factor = dataloader_prefetch_factor,
remove_unused_columns = remove_unused_columns,
label_names = label_names,
train_sampling_strategy = train_sampling_strategy,
length_column_name = length_column_name,
ddp_find_unused_parameters = ddp_find_unused_parameters,
ddp_bucket_cap_mb = ddp_bucket_cap_mb,
ddp_broadcast_buffers = ddp_broadcast_buffers,
ddp_backend = ddp_backend,
ddp_timeout = ddp_timeout,
fsdp = fsdp,
fsdp_config = fsdp_config,
deepspeed = deepspeed,
debug = debug,
skip_memory_metrics = skip_memory_metrics,
do_train = do_train,
do_eval = do_eval,
do_predict = do_predict,
resume_from_checkpoint = resume_from_checkpoint,
warmup_ratio = warmup_ratio,
logging_dir = logging_dir,
local_rank = local_rank,
model_init_kwargs = model_init_kwargs,
disable_dropout = disable_dropout,
max_prompt_length = max_prompt_length,
num_generations = num_generations,
max_completion_length = max_completion_length,
ds3_gather_for_generation = ds3_gather_for_generation,
shuffle_dataset = shuffle_dataset,
generation_batch_size = generation_batch_size,
steps_per_generation = steps_per_generation,
temperature = temperature,
top_p = top_p,
top_k = top_k,
min_p = min_p,
generation_kwargs = generation_kwargs,
repetition_penalty = repetition_penalty,
use_transformers_paged = use_transformers_paged,
cache_implementation = cache_implementation,
use_vllm = use_vllm,
vllm_mode = vllm_mode,
vllm_model_impl = vllm_model_impl,
vllm_enable_sleep_mode = vllm_enable_sleep_mode,
vllm_guided_decoding_regex = vllm_guided_decoding_regex,
vllm_server_base_url = vllm_server_base_url,
vllm_server_host = vllm_server_host,
vllm_server_port = vllm_server_port,
vllm_server_timeout = vllm_server_timeout,
vllm_gpu_memory_utilization = vllm_gpu_memory_utilization,
vllm_tensor_parallel_size = vllm_tensor_parallel_size,
beta = beta,
num_iterations = num_iterations,
epsilon = epsilon,
delta = delta,
epsilon_high = epsilon_high,
importance_sampling_level = importance_sampling_level,
reward_weights = reward_weights,
scale_rewards = scale_rewards,
loss_type = loss_type,
mask_truncated_completions = mask_truncated_completions,
sync_ref_model = sync_ref_model,
ref_model_mixup_alpha = ref_model_mixup_alpha,
ref_model_sync_steps = ref_model_sync_steps,
top_entropy_quantile = top_entropy_quantile,
use_liger_loss = use_liger_loss,
vllm_importance_sampling_correction = vllm_importance_sampling_correction,
vllm_importance_sampling_cap = vllm_importance_sampling_cap,
log_completions = log_completions,
num_completions_to_print = num_completions_to_print,
wandb_log_unique_prompts = wandb_log_unique_prompts,**kwargs)
self.vllm_sampling_params = vllm_sampling_params
self.unsloth_num_chunks = unsloth_num_chunks
if unsloth_grpo_mini_batch is not None:
if self.generation_batch_size >= unsloth_grpo_mini_batch:
self.unsloth_grpo_mini_batch = unsloth_grpo_mini_batch
else:
raise ValueError(
f"Unsloth GRPO mini batch size needs to be less than or equal to the effective generation batch size, "
f"which is self.per_device_train_batch_size * gradient_accumulation_steps."
)
self.unsloth_logit_chunk_multiplier = unsloth_logit_chunk_multiplier
pass
class _UnslothGRPOTrainer(BaseTrainer):
""""""
_tag_names = ["trl", "grpo"]
_name = "GRPO"
_paper = {
"title": "DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models",
"id": "2402.03300",
# docstyle-ignore
"citation": textwrap.dedent("""\
@article{shao2024deepseekmath,
title = {{DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models}},
author = {Zhihong Shao and Peiyi Wang and Qihao Zhu and Runxin Xu and Junxiao Song and Mingchuan Zhang and Y. K. Li and Y. Wu and Daya Guo},
year = 2024,
eprint = {arXiv:2402.03300},
}
"""),
}
def __init__(
self,
model: Union[str, PreTrainedModel],
reward_funcs: Union[RewardFunc, list[RewardFunc]],
args: Optional[GRPOConfig] = None,
train_dataset: Optional[Union[Dataset, IterableDataset]] = None,
eval_dataset: Optional[Union[Dataset, IterableDataset, dict[str, Union[Dataset, IterableDataset]]]] = None,
processing_class: Optional[Union[PreTrainedTokenizerBase, ProcessorMixin]] = None,
reward_processing_classes: Optional[Union[PreTrainedTokenizerBase, list[PreTrainedTokenizerBase]]] = None,
callbacks: Optional[list[TrainerCallback]] = None,
optimizers: tuple[Optional[torch.optim.Optimizer], Optional[torch.optim.lr_scheduler.LambdaLR]] = (None, None),
peft_config: Optional["PeftConfig"] = None,
):
if hasattr(model, 'vllm_engine') and hasattr(args, 'use_vllm'):
if (getattr(args, 'use_vllm', False) == False):
args.use_vllm = True
args.vllm_mode='colocate'
if os.environ.get('UNSLOTH_VLLM_STANDBY', '0') == '1':
args.vllm_enable_sleep_mode=True
# Args
if args is None:
model_name = model if isinstance(model, str) else model.config._name_or_path
model_name = model_name.split("/")[-1]
args = GRPOConfig(f"{model_name}-GRPO")
# Models
# Trained model
model_init_kwargs = args.model_init_kwargs or {}
if isinstance(model, str):
model_id = model
dtype = model_init_kwargs.get("dtype")
if isinstance(dtype, torch.dtype) or dtype == "auto" or dtype is None:
pass # dtype is already a torch.dtype or "auto" or None
elif isinstance(dtype, str): # it's a str, but not "auto"
dtype = getattr(torch, dtype)
model_init_kwargs["dtype"] = dtype
else:
raise ValueError(
"Invalid `dtype` passed to `GRPOConfig`. Expected either 'auto' or a string representing "
f"a `torch.dtype` (e.g., 'float32'), but got {dtype}."
)
# Disable caching if gradient checkpointing is enabled [not supported]
config = AutoConfig.from_pretrained(model_id)
architecture = getattr(transformers, config.architectures[0])
model = architecture.from_pretrained(model_id, **model_init_kwargs)
else:
model_id = model.config._name_or_path
if args.model_init_kwargs is not None:
logger.warning(
"You passed `model_init_kwargs` to the `GRPOConfig`, but your model is already instantiated. "
"The `model_init_kwargs` will be ignored."
)
# Some models [SmolVLM/Idefics3] don't support `logits_to_keep` argument and error out if we pass it
# Inspect the forward method before we wrap the model with PEFT
self.model_kwarg_keys = (
inspect.signature(model.forward).parameters.keys()
if not hasattr(model, "get_base_model")
else inspect.signature(model.get_base_model().forward).parameters.keys()
)
if False:
pass
# Processing class
if processing_class is None:
processing_class = AutoProcessor.from_pretrained(model.config._name_or_path, truncation_side="left")
# Handle pad token for processors or tokenizers
if isinstance(processing_class, ProcessorMixin):
tokenizer = processing_class.tokenizer
elif isinstance(processing_class, PreTrainedTokenizerBase):
tokenizer = processing_class
else:
raise TypeError("The `processing_class` must be either a `PreTrainedTokenizerBase` or a `ProcessorMixin`")
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
self.pad_token = tokenizer.pad_token
self.pad_token_id = tokenizer.pad_token_id
self.eos_token_id = tokenizer.eos_token_id
# Reward functions
if not isinstance(reward_funcs, list):
reward_funcs = [reward_funcs]
self.reward_func_names = []
for i, reward_func in enumerate(reward_funcs):
if isinstance(reward_func, str):
reward_funcs[i] = AutoModelForSequenceClassification.from_pretrained(
reward_func, num_labels=1, **model_init_kwargs
)
if isinstance(reward_funcs[i], nn.Module): # Use Module over PretrainedModel for compat w/ compiled models
self.reward_func_names.append(reward_funcs[i].config._name_or_path.split("/")[-1])
else:
self.reward_func_names.append(reward_funcs[i].__name__)
self.reward_funcs = reward_funcs
# Reward weights
if args.reward_weights is not None:
if len(args.reward_weights) != len(reward_funcs):
raise ValueError(
f"Number of reward weights ({len(args.reward_weights)}) must match number of reward "
f"functions ({len(reward_funcs)})"
)
self.reward_weights = torch.tensor(args.reward_weights, dtype=torch.float32)
else:
self.reward_weights = torch.ones(len(reward_funcs), dtype=torch.float32)
# Reward processing class
if reward_processing_classes is None:
reward_processing_classes = [None] * len(reward_funcs)
elif not isinstance(reward_processing_classes, list):
reward_processing_classes = [reward_processing_classes]
if len(reward_processing_classes) != len(reward_funcs):
raise ValueError(
f"The number of reward processing classes ({len(reward_processing_classes)}) must match the number of "
f"reward functions ({len(reward_funcs)})."
)
for i, (reward_processing_class, reward_func) in enumerate(zip(reward_processing_classes, reward_funcs)):
if isinstance(reward_func, PreTrainedModel):
if reward_processing_class is None:
reward_processing_class = AutoTokenizer.from_pretrained(reward_func.config._name_or_path)
if reward_processing_class.pad_token_id is None:
reward_processing_class.pad_token = reward_processing_class.eos_token
# The reward model computes the reward for the latest non-padded token in the input sequence.
# So it's important to set the pad token ID to the padding token ID of the processing class.
reward_func.config.pad_token_id = reward_processing_class.pad_token_id
reward_processing_classes[i] = reward_processing_class
self.reward_processing_classes = reward_processing_classes
# Training arguments
self.max_prompt_length = args.max_prompt_length
self.max_completion_length = args.max_completion_length # = |o_i| in the GRPO paper
self.num_generations = args.num_generations # = G in the GRPO paper
self.temperature = args.temperature
self.top_p = args.top_p
self.top_k = args.top_k
self.min_p = args.min_p
self.repetition_penalty = args.repetition_penalty
self.use_transformers_paged = args.use_transformers_paged
self.use_vllm = args.use_vllm
self.vllm_mode = args.vllm_mode
self.vllm_gpu_memory_utilization = args.vllm_gpu_memory_utilization # only applies to colocation mode
self.vllm_tensor_parallel_size = args.vllm_tensor_parallel_size # only applies to colocation mode
self.vllm_importance_sampling_correction = args.vllm_importance_sampling_correction
self.vllm_importance_sampling_cap = args.vllm_importance_sampling_cap
self.use_liger_loss = args.use_liger_loss
self.loss_type = args.loss_type
self.scale_rewards = args.scale_rewards
self.importance_sampling_level = args.importance_sampling_level
self.mask_truncated_completions = args.mask_truncated_completions
self.top_entropy_quantile = args.top_entropy_quantile
if self.use_liger_loss and self.top_entropy_quantile < 1.0:
raise NotImplementedError(
"Liger Kernels don't currently support masking token positions based on entropy."
)
if self.use_liger_loss and not self.importance_sampling_level == "token":
raise NotImplementedError(
"Liger Kernels currently only support token-level importance sampling. Please set"
"`importance_sampling_level` to 'token'."
)
# Datasets
self.shuffle_dataset = args.shuffle_dataset
if (
isinstance(train_dataset, IterableDataset)
or isinstance(eval_dataset, IterableDataset)
or (
isinstance(eval_dataset, dict) and any(isinstance(ds, IterableDataset) for ds in eval_dataset.values())
)
):
# See https://github.com/huggingface/trl/issues/3213
raise NotImplementedError(
"Iterable datasets are not yet supported in GRPOTrainer. Please use a standard dataset instead."
)
# Multi-step
self.num_iterations = args.num_iterations # = 𝜇 in the GRPO paper
self.epsilon_low = args.epsilon
self.epsilon_high = args.epsilon_high if args.epsilon_high is not None else args.epsilon
# Tracks the number of iterations [forward + backward passes], including those within a grad accum cycle
self._step = 0
# Buffer the batch to reuse generated outputs across multiple updates. For more details, see
# `_get_train_sampler` and `_prepare_inputs`.
self._buffered_inputs = None
# The trainer estimates the number of FLOPs [floating-point operations] using the number of elements in the
# input tensor associated with the key "input_ids". However, in GRPO, the sampled data does not include the
# "input_ids" key. Instead, the available keys is "prompt". As a result, the trainer issues the warning:
# "Could not estimate the number of tokens of the input, floating-point operations will not be computed." To
# suppress this warning, we set the "estimate_tokens" key in the model's "warnings_issued" dictionary to True.
# This acts as a flag to indicate that the warning has already been issued.
model.warnings_issued["estimate_tokens"] = True
super().__init__(
model=model,
args=args,
data_collator=identity, # No data collation is needed in GRPO
train_dataset=train_dataset,
eval_dataset=eval_dataset,
processing_class=processing_class,
callbacks=callbacks,
optimizers=optimizers,
# In Trainer, `training_step` scales the loss by `gradient_accumulation_steps` only if `compute_loss_func`
# is None. For DAPO, loss scaling instead depends on the total number of completions tokens across the
# global accumulated batch. To control scaling ourselves, we must disable Trainer’s built-in scaling. The
# simplest [though a bit hacky] way is to set `compute_loss_func` to any non-None value, which bypasses
# that behavior without rewriting `training_step`.
compute_loss_func="non-None value to disable scaling",
)
# Reference model
self.beta = args.beta
if self.beta == 0.0:
# If beta is 0.0, the reference model is not needed
self.ref_model = None
elif is_peft_model(model):
# If PEFT is used, the reference model is not needed since the adapter can be disabled
# to revert to the initial model.
self.ref_model = None
else:
# For deepspeed, fsdp or non-distributed models, create a reference model from scratch
config = AutoConfig.from_pretrained(model_id)
architecture = getattr(transformers, config.architectures[0])
self.ref_model = architecture.from_pretrained(model_id, **model_init_kwargs)
# Disable dropout in the models
if args.disable_dropout:
disable_dropout_in_model(model)
if self.ref_model is not None:
disable_dropout_in_model(self.ref_model)
# Liger loss
if self.use_liger_loss:
if not is_liger_kernel_available():
raise ImportError(
"Liger is required to use `liger_loss` as the GRPO loss. Run `pip install liger-kernel`."
)
# redirect the model.module forward to the model forward to ensure pre-forward hooks are called
self._forward_redirection = _ForwardRedirection()
self.liger_grpo_loss = LigerFusedLinearGRPOLoss(
beta=self.beta,
epsilon_low=self.epsilon_low,
epsilon_high=self.epsilon_high,
temperature=self.temperature,
use_ref_model=self.beta != 0.0,
loss_type=self.loss_type,
max_completion_length=self.max_completion_length,
)
# Initialize the metrics
self._metrics = {"train": defaultdict(list), "eval": defaultdict(list)}
self._total_train_tokens = 0
self.log_completions = args.log_completions
self.wandb_log_unique_prompts = args.wandb_log_unique_prompts
self.num_completions_to_print = args.num_completions_to_print
# Keep logs sized to the generation batch to record only outputs from the latest model update.
self._logs = {
"images": deque(maxlen=args.generation_batch_size),
"prompt": deque(maxlen=args.generation_batch_size),
"completion": deque(maxlen=args.generation_batch_size),
"rewards": defaultdict(lambda: deque(maxlen=args.generation_batch_size)),
"advantages": deque(maxlen=args.generation_batch_size),
}
# Ensure each process receives a unique seed to prevent duplicate completions when generating with
# transformers if num_generations exceeds per_device_train_batch_size. We could skip it if we use vLLM, but
# it's safer to set it in all cases.
set_seed(args.seed, device_specific=True)
if self.use_vllm:
if not is_vllm_available():
raise ImportError(
"vLLM is not available and `use_vllm` is set to True. Please install vLLM with "
"`pip install trl[vllm]` to use it."
)
if self.vllm_mode == "server":
if self.accelerator.is_main_process:
if args.vllm_server_base_url is not None:
base_url = args.vllm_server_base_url
else:
base_url = f"http://{args.vllm_server_host}:{args.vllm_server_port}"
self.vllm_client = VLLMClient(base_url=base_url, connection_timeout=args.vllm_server_timeout)
self.vllm_client.init_communicator(device=torch.cuda.current_device())
elif self.vllm_mode == "colocate":
if not self.accelerator.num_processes % self.vllm_tensor_parallel_size == 0:
raise ValueError(
f"vllm_tensor_parallel_size ({self.vllm_tensor_parallel_size}) must divide world size "
f"({self.accelerator.num_processes}) evenly."
)
if self.vllm_tensor_parallel_size > 1:
self.tp_group, _ = torch.distributed.new_subgroups_by_enumeration(
[
list(range(i * self.vllm_tensor_parallel_size, (i + 1) * self.vllm_tensor_parallel_size))
for i in range(self.accelerator.num_processes // self.vllm_tensor_parallel_size)
]
)
os.environ["RANK"] = str(self.accelerator.process_index)
os.environ["LOCAL_RANK"] = str(self.accelerator.local_process_index)
os.environ["WORLD_SIZE"] = str(self.accelerator.num_processes)
ensure_master_addr_port()
if self.max_prompt_length is not None and self.max_completion_length is not None:
max_model_len = self.max_prompt_length + self.max_completion_length
else:
max_model_len = None
if getattr(getattr(model, 'vllm_engine', None), 'shared_weights', False):
self.llm = model.vllm_engine
else:
self.llm = LLM(
model=model.name_or_path,
tensor_parallel_size=args.vllm_tensor_parallel_size,
gpu_memory_utilization=self.vllm_gpu_memory_utilization,
max_num_seqs=self.args.per_device_train_batch_size
* self.vllm_tensor_parallel_size
* self.args.steps_per_generation,
max_model_len=max_model_len,
distributed_executor_backend="external_launcher",
seed=self.accelerator.process_index // self.vllm_tensor_parallel_size,
max_num_batched_tokens=4096,
model_impl=self.args.vllm_model_impl,
enable_sleep_mode=self.args.vllm_enable_sleep_mode,
logprobs_mode="processed_logprobs",
)
if self.args.vllm_enable_sleep_mode:
self.llm.sleep(level=1)
else:
raise ValueError(f"vllm_mode must be either 'server' or 'colocate', got '{self.vllm_mode}'.")
self.guided_decoding_regex = args.vllm_guided_decoding_regex
self._last_loaded_step = -1
self.accelerator.wait_for_everyone()
else:
generation_kwargs = {
"max_new_tokens": self.max_completion_length,
"do_sample": True,
"pad_token_id": tokenizer.pad_token_id,
"bos_token_id": tokenizer.bos_token_id,
"eos_token_id": tokenizer.eos_token_id,
"temperature": self.temperature,
"top_p": self.top_p,
"top_k": self.top_k,
"min_p": self.min_p,
"repetition_penalty": self.repetition_penalty,
"cache_implementation": args.cache_implementation,
}
if args.generation_kwargs is not None:
generation_kwargs.update(args.generation_kwargs)
self.generation_config = GenerationConfig(**generation_kwargs)
# Gradient accumulation requires scaled loss. Normally, loss scaling in the parent class depends on whether the
# model accepts loss-related kwargs. Since we compute our own loss, this check is irrelevant. We set
# self.model_accepts_loss_kwargs to False to enable scaling.
self.model_accepts_loss_kwargs = False
# Add tags to the model
self.model.add_model_tags(self._tag_names)
if self.ref_model is not None:
if self.is_deepspeed_enabled:
self.ref_model = prepare_deepspeed(self.ref_model, self.accelerator)
elif self.is_fsdp_enabled:
self.ref_model = prepare_fsdp(self.ref_model, self.accelerator)
else:
self.ref_model = self.accelerator.prepare_model(self.ref_model, evaluation_mode=True)
if args.sync_ref_model:
self.add_callback(SyncRefModelCallback(ref_model=self.ref_model, accelerator=self.accelerator))
for i, reward_func in enumerate(self.reward_funcs):
if isinstance(reward_func, PreTrainedModel):
if self.is_deepspeed_enabled:
self.reward_funcs[i] = prepare_deepspeed(reward_func, self.accelerator)
else:
# set device placement to True to make `prepare_model` move `reward_func` to device when using fsdp
self.reward_funcs[i] = self.accelerator.prepare_model(
reward_func, evaluation_mode=True, device_placement=True
)
def _set_signature_columns_if_needed(self):
# If `self.args.remove_unused_columns` is True, non-signature columns are removed.
# By default, this method sets `self._signature_columns` to the model's expected inputs.
# In GRPOTrainer, we preprocess data, so using the model's signature columns doesn't work.
# Instead, we set them to the columns expected by the `training_step` method, hence the override.
if self._signature_columns is None:
self._signature_columns = ["prompt", "image", "images"]
# This method overrides `Trainer.get_train_dataloader` to support our custom batching strategy.
# Instead of returning a standard per-step batch (i.e., `per_device_batch_size), our dataloader loads an
# *generation* batch (i.e., `per_device_batch_size × steps_per_generation`). This allows us to generate completions
# once every steps_per_generation step—rather than once per accumulation step—which is significantly more
# efficient. The only change from the original implementation is multiplying the batch size by
# `steps_per_generation`. Thus, `_prepare_inputs` is called with this *generation* batch, and it handles the
# splitting internally.
# Maintenance note: This method is a copy-paste of the original `Trainer.get_train_dataloader` with only one line
# modification. As a result, some parts of the method aren't relevant to GRPO, but we keep them to stay one line
# apart from the super method, ensuring easier maintenance in the future.
def get_train_dataloader(self):
if self.train_dataset is None:
raise ValueError("Trainer: training requires a train_dataset.")
train_dataset = self.train_dataset
data_collator = self.data_collator
if is_datasets_available() and isinstance(train_dataset, datasets.Dataset):
train_dataset = self._remove_unused_columns(train_dataset, description="training")
else:
data_collator = self._get_collator_with_removed_columns(data_collator, description="training")
dataloader_params = {
"batch_size": self._train_batch_size * self.args.steps_per_generation, # < this is the change
"collate_fn": data_collator,
"num_workers": self.args.dataloader_num_workers,
"pin_memory": self.args.dataloader_pin_memory,
"persistent_workers": self.args.dataloader_persistent_workers,
}
if not isinstance(train_dataset, torch.utils.data.IterableDataset):
dataloader_params["sampler"] = self._get_train_sampler()
dataloader_params["drop_last"] = self.args.dataloader_drop_last
dataloader_params["worker_init_fn"] = partial(
seed_worker, num_workers=self.args.dataloader_num_workers, rank=self.args.process_index
)
dataloader_params["prefetch_factor"] = self.args.dataloader_prefetch_factor
return self.accelerator.prepare(DataLoader(train_dataset, **dataloader_params))
def _get_train_sampler(self, dataset: Optional[Dataset] = None) -> Sampler:
# Returns a sampler that
# 1. ensures each prompt is repeated across multiple processes. This guarantees that identical prompts are
# distributed to different GPUs, allowing rewards to be computed and normalized correctly within each prompt
# group. Using the same seed across processes ensures consistent prompt assignment, preventing discrepancies
# in group formation.
# 2. repeats the batch multiple times to allow reusing generations across multiple updates. Refer to
# _prepare_inputs to see how the generations are stored and reused.
# In the following figure, the values are the prompt indices. The first row shows the first sampled batch, the
# second row shows the second sampled batch, and so on.
#
# | GPU 0 | GPU 1 |
#
# global_step step <-───> num_generations=2
# <-───────> per_device_train_batch_size=3
# grad_accum ▲ ▲ 0 0 0 0 1 1 2 2 <- Generate for the first `steps_per_generation` (prompts 0 to 11); store the completions; use the first slice to compute the loss
# =2 ▼ | 0 1 3 3 4 4 5 5 <- Take the stored generations and use the second slice to compute the loss
# |
# | 1 2 6 6 7 7 8 8 <- Take the stored generations and use the third slice to compute the loss
# steps_per_gen=4 ▼ 1 3 9 9 10 10 11 11 <- Take the stored generations and use the fourth slice to compute the loss
#
# 2 4 12 12 13 13 14 14 <- Generate for the second `steps_per_generation` (prompts 12 to 23); store the completions; use the first slice to compute the loss
# 2 5 15 15 16 16 17 17 <- Take the stored generations and use the second slice to compute the loss
# ...
if dataset is None:
dataset = self.train_dataset
return RepeatSampler(
data_source=dataset,
mini_repeat_count=self.num_generations,
batch_size=self.args.generation_batch_size // self.num_generations,
repeat_count=self.num_iterations * self.args.steps_per_generation,
shuffle=self.shuffle_dataset,
seed=self.args.seed,
)
def _get_eval_sampler(self, eval_dataset) -> Sampler:
# See _get_train_sampler for an explanation of the sampler.
return RepeatSampler(
data_source=eval_dataset,
mini_repeat_count=self.num_generations,
seed=self.args.seed,
)
@profiling_decorator
def _get_last_hidden_state(
self,
unwrapped_model,
input_ids,
attention_mask,
logits_to_keep,
pixel_values=None,
image_grid_thw=None,
pixel_attention_mask=None,
image_sizes=None,
):
if is_peft_model(unwrapped_model):
unwrapped_model = unwrapped_model.base_model.model
# Build model inputs - check if the model supports logits_to_keep (some models and VLMs don't)
model_inputs = {"input_ids": input_ids, "attention_mask": attention_mask}
# For Qwen models:
if image_grid_thw is not None and pixel_values is not None:
model_inputs["image_grid_thw"] = image_grid_thw
# For Gemma, SmolVLM2, LLaVa-Next etc.:
if pixel_values is not None:
model_inputs["pixel_values"] = pixel_values
# For SmolVLM2
if pixel_attention_mask is not None:
model_inputs["pixel_attention_mask"] = pixel_attention_mask
# For LLaVa-Next
if image_sizes is not None:
model_inputs["image_sizes"] = image_sizes
# Only add logits_to_keep if the model supports it
if "logits_to_keep" in self.model_kwarg_keys:
# We add 1 to `logits_to_keep` because the last logits of the sequence is later excluded
model_inputs["logits_to_keep"] = logits_to_keep + 1
model_inputs["use_cache"] = False # only used in generation; set False to suppress warnings
last_hidden_state = unwrapped_model.model(**model_inputs).last_hidden_state
# Exclude the last value: it corresponds to the next token pred
last_hidden_state = last_hidden_state[:, :-1, :] # (B, L-1, H)
# Only keep the last logits_to_keep. For model that support logits_to_keep, this is a no-op.
last_hidden_state = last_hidden_state[:, -logits_to_keep:, :] # (B, logits_to_keep, H)
return last_hidden_state
def get_high_entropy_mask(self, entropies: torch.Tensor, mask: torch.Tensor, threshold: float) -> torch.Tensor:
"""
Returns a binary mask identifying tokens whose entropy exceeds a given quantile threshold.
Args:
entropies (`torch.Tensor`):
Tensor of shape (batch_size, seq_len) with per-token entropy values.
mask (`torch.Tensor`):
Binary mask of the same shape as `entropies`, where `1` indicates valid tokens and `0` padding.
threshold (`float`):
Quantile threshold between `0.0` and `1.0` to select high-entropy tokens.
Returns:
`torch.Tensor`:
Boolean mask of shape (batch_size, seq_len), where `True` indicates tokens with entropy >= threshold
and `False` otherwise.
"""
local = entropies[mask.bool()].float()
# Use a negative pad_value as a sentinel because entropy values are always >= 0.
# This guarantees that the sentinel cannot collide with any real entropy value.
pad_value = -1e9
# Pad across processes so that every rank has the same tensor length
padded = self.accelerator.pad_across_processes(local, dim=0, pad_index=pad_value)
gathered = self.accelerator.gather(padded)
# Drop sentinel values (safe because no entropy can be negative)
gathered = gathered[gathered != pad_value]
if gathered.numel() == 0:
return torch.zeros_like(entropies, dtype=torch.bool)
entropy_threshold = torch.quantile(gathered, threshold)
masked_entropies = entropies * mask.float()
entropy_mask = masked_entropies >= entropy_threshold
return entropy_mask & mask.bool() # ensure padding tokens are always masked out
def _get_per_token_logps_and_entropies(
self,
model,
input_ids,
attention_mask,
logits_to_keep,
batch_size = None,
compute_entropy = False,
compute_efficient = False,
*args,
**kwargs,
):
# All Unsloth code here in this function is licensed under AGPL3
# if True: # os.environ.get('UNSLOTH_USE_NEW_MODEL', '0') == '0':
# return None, None # logps, entropies Unsloth efficient GRPO
if compute_efficient:
return None, None
else:
if not hasattr(self, "_autocast_dtype"):
self._autocast_dtype = (
torch.float16
if os.environ.get("ACCELERATE_MIXED_PRECISION", "fp16") == "fp16"
else torch.bfloat16
)
if os.environ.get("UNSLOTH_FORCE_FLOAT32", "0") == "1":
self._autocast_dtype = torch.float16
pixel_values, image_grid_thw = (
kwargs.get("pixel_values", None),
kwargs.get("image_grid_thw", None),
)
pixel_attention_mask, image_sizes = (
kwargs.get("pixel_attention_mask", None),
kwargs.get("image_sizes", None),
)
num_images = kwargs.get("num_images", None)
# Transformers 5.x needs token_type_ids/mm_token_type_ids for some vision models
token_type_ids = kwargs.get("token_type_ids", None)
mm_token_type_ids = kwargs.get("mm_token_type_ids", None)
if mm_token_type_ids is not None or image_grid_thw is not None:
mm_token_type_ids = _unsloth_fix_mm_token_type_ids(
self.processing_class, input_ids, mm_token_type_ids
)
unwrapped_model = self.accelerator.unwrap_model(
model, keep_fp32_wrapper = False
)
lm_head = self.model.get_output_embeddings().weight
dtype_bytes = (
16 if self._autocast_dtype in [torch.float16, torch.bfloat16] else 32
)
total_rows = input_ids.shape[0]
seq_len = input_ids.shape[1]
hidden_dim = lm_head.shape[1]
vocab_dim = lm_head.shape[0]
if self.args.unsloth_grpo_mini_batch is None:
B, multiplier = autotune_batch_and_chunks(
total_rows,
seq_len,
hidden_dim,
vocab_dim,
dtype_bytes,
self.args.unsloth_logit_chunk_multiplier,
)
B = total_rows // B
else:
B = self.args.unsloth_grpo_mini_batch
if self.args.unsloth_logit_chunk_multiplier is None:
multiplier = max(4, seq_len // 4096)
else:
multiplier = self.args.unsloth_logit_chunk_multiplier
all_logprobs_list = []
if pixel_values is None:
left_pad_tokens_per_prompt = calculate_pad_tokens_in_prompt(
input_ids, logits_to_keep, self.processing_class.pad_token_id
)
max_left_pad = torch.max(left_pad_tokens_per_prompt).item()
input_ids = left_pack_padding(
input_ids, self.processing_class.pad_token_id
)
attention_mask = input_ids != self.processing_class.pad_token_id
attention_mask = attention_mask.to(attention_mask.dtype)
else:
max_left_pad = 0
def slice_sample_axis(value, start, end):
if value is None:
return None
return value[start:end]
import math
total_samples = input_ids.shape[0]
batch_size = math.ceil(total_samples / B)
if isinstance(num_images, torch.Tensor):
num_images = num_images.detach().cpu().reshape(-1).tolist()
if (
image_grid_thw is not None
and pixel_values is not None
and num_images is not None
):
rows_per_image = image_grid_thw.prod(dim = -1)
rows_per_sample = torch.split(rows_per_image, num_images)
rows_per_sample = torch.stack([s.sum() for s in rows_per_sample])
# why: cum_rows is indexed via .item() inside the per-chunk loop;
# keeping it on CPU avoids per-iteration GPU->CPU sync.
cum_rows = torch.cat(
[
torch.tensor([0], device = rows_per_sample.device),
rows_per_sample.cumsum(0),
]
).cpu()
cum_imgs = torch.tensor([0] + num_images).cumsum(0)
else:
cum_rows = None
cum_imgs = None
def _first_dim_len(value):
if value is None:
return None
if hasattr(value, "shape"):
return value.shape[0]
try:
return len(value)
except TypeError:
return None
total_images = sum(num_images) if num_images is not None else None
_image_sizes_n = _first_dim_len(image_sizes)
input_ids_chunks = []
attention_mask_chunks = []
pixel_values_chunks = []
image_grid_thw_chunks = []
pixel_attention_mask_chunks = []
image_sizes_chunks = []
token_type_ids_chunks = []
mm_token_type_ids_chunks = []
current_pixel_idx = 0
# TRL 0.23.0 batching logic
for start in range(0, total_samples, batch_size):
end = min(start + batch_size, total_samples)
input_ids_chunks.append(input_ids[start:end])
attention_mask_chunks.append(attention_mask[start:end])
token_type_ids_chunks.append(
slice_sample_axis(token_type_ids, start, end)
)
mm_token_type_ids_chunks.append(
slice_sample_axis(mm_token_type_ids, start, end)
)
if image_grid_thw is not None and pixel_values is not None:
if num_images is None:
grid_slice = image_grid_thw[start:end]
batch_pixel_count = grid_slice.prod(dim = -1).sum().item()
start_pixel_idx = current_pixel_idx
end_pixel_idx = current_pixel_idx + batch_pixel_count
current_pixel_idx = end_pixel_idx
img_start = img_end = None
else:
start_pixel_idx = cum_rows[start].item()
end_pixel_idx = cum_rows[end].item()
img_start = cum_imgs[start].item()
img_end = cum_imgs[end].item()
grid_slice = image_grid_thw[img_start:img_end]
image_grid_thw_chunks.append(grid_slice)
pixel_values_chunks.append(
pixel_values[start_pixel_idx:end_pixel_idx]
)
if image_sizes is None:
image_sizes_chunks.append(None)
elif (
num_images is not None
and _image_sizes_n == total_images
and img_start is not None
):
image_sizes_chunks.append(image_sizes[img_start:img_end])
else:
image_sizes_chunks.append(
slice_sample_axis(image_sizes, start, end)
)
if pixel_attention_mask is None:
pixel_attention_mask_chunks.append(None)
elif (
num_images is not None
and img_start is not None
and pixel_attention_mask.shape[0] == image_grid_thw.shape[0]
):
pixel_attention_mask_chunks.append(
pixel_attention_mask[img_start:img_end]
)
elif (
pixel_attention_mask.shape[0] == pixel_values.shape[0]
and pixel_attention_mask.shape[0] != input_ids.shape[0]
):
pixel_attention_mask_chunks.append(
pixel_attention_mask[start_pixel_idx:end_pixel_idx]
)
else:
pixel_attention_mask_chunks.append(
pixel_attention_mask[start:end]
)
else:
pixel_values_chunks.append(None)
image_grid_thw_chunks.append(None)
pixel_attention_mask_chunks.append(None)
image_sizes_chunks.append(
slice_sample_axis(image_sizes, start, end)
)
temperature = self.temperature
logit_softcapping = _unsloth_get_final_logit_softcapping(model.config)
logit_scale_multiply = getattr(model.config, "logit_scale", 0)
if logit_scale_multiply is None:
logit_scale_multiply = 0
logit_scale_divide = getattr(model.config, "logits_scaling", 0)
if logit_scale_divide is None:
logit_scale_divide = 0
zipped_inputs = zip(
input_ids_chunks,
attention_mask_chunks,
pixel_values_chunks,
image_grid_thw_chunks,
pixel_attention_mask_chunks,
image_sizes_chunks,
token_type_ids_chunks,
mm_token_type_ids_chunks,
)
os.environ["UNSLOTH_RETURN_HIDDEN_STATES"] = "1"
with _get_inference_mode_context_manager(model):
for (
input_ids_chunk,
attention_mask_chunk,
pixel_values_chunk,
image_grid_thw_chunk,
pixel_attention_mask_chunk,
image_sizes_chunk,
token_type_ids_chunk,
mm_token_type_ids_chunk,
) in zipped_inputs:
_extra_vision_kwargs = {}
if token_type_ids_chunk is not None:
_extra_vision_kwargs["token_type_ids"] = token_type_ids_chunk
if mm_token_type_ids_chunk is not None:
_extra_vision_kwargs["mm_token_type_ids"] = (
mm_token_type_ids_chunk
)
with torch.amp.autocast(
device_type = "cuda", dtype = self._autocast_dtype
):
if pixel_values is None:
logits_chunk = unwrapped_model(
input_ids = input_ids_chunk,
attention_mask = attention_mask_chunk,
pixel_values = pixel_values_chunk,
image_grid_thw = image_grid_thw_chunk,
pixel_attention_mask = pixel_attention_mask_chunk,
image_sizes = image_sizes_chunk,
**_extra_vision_kwargs,
).logits
completion_input_ids_chunk = input_ids_chunk[
:, -(logits_to_keep + max_left_pad) :
]
logits_chunk = logits_chunk[
:, -(logits_to_keep + max_left_pad + 1) :, :
]
logits_chunk = logits_chunk[:, :-1, :]
logprobs_chunk = (
chunked_hidden_states_selective_log_softmax(
logits_chunk,
lm_head,
completion_input_ids_chunk,
chunks = input_ids_chunk.shape[0] * multiplier,
logit_scale_multiply = logit_scale_multiply,
logit_scale_divide = logit_scale_divide,
logit_softcapping = logit_softcapping,
temperature = temperature,
)
)
else:
# Essentially, for VLMs we do not go via the optimized path in models/,
# so we don't encounter the Flash Attn left-padding issue.
logits_chunk = unwrapped_model(
input_ids = input_ids_chunk,
attention_mask = attention_mask_chunk,
pixel_values = pixel_values_chunk,
image_grid_thw = image_grid_thw_chunk,
pixel_attention_mask = pixel_attention_mask_chunk,
image_sizes = image_sizes_chunk,
logits_to_keep = logits_to_keep + 1,
**_extra_vision_kwargs,
).logits
logits_chunk = logits_chunk[:, :-1, :]
completion_input_ids_chunk = input_ids_chunk[
:, -logits_to_keep:
]
# Guard: check if model returned hidden states or logits
if logits_chunk.shape[-1] == lm_head.shape[1]:
logprobs_chunk = (
chunked_hidden_states_selective_log_softmax(
logits_chunk,
lm_head,
completion_input_ids_chunk,
chunks = input_ids_chunk.shape[0] * multiplier,
logit_scale_multiply = logit_scale_multiply,
logit_scale_divide = logit_scale_divide,
logit_softcapping = logit_softcapping,
temperature = temperature,
)
)
else:
# Model returned logits directly - scaling/softcapping already applied by model forward
logprobs_chunk = chunked_selective_log_softmax(
logits_chunk,
completion_input_ids_chunk,
temperature,
)
# This is needed to avoid race conditions with GPT OSS offload_embbed=True
# However, it seems that this line does not slow down or disrupt models.
device_synchronize()
all_logprobs_list.append(logprobs_chunk)
logprobs = torch.cat(all_logprobs_list, dim = 0)
entropies = None
os.environ["UNSLOTH_RETURN_HIDDEN_STATES"] = "0"
return logprobs.detach(), entropies # logps, entropies
# input_ids = input_ids[:, -logits_to_keep:]
# For transformers<=4.48, logits_to_keep argument isn't supported, so here we drop logits ourselves.
# See https://github.com/huggingface/trl/issues/2770
# logits = logits[:, -logits_to_keep:]
# return logits
# See https://huggingface.co/blog/the_n_implementation_details_of_rlhf_with_ppo#policy-training-implementation-details
# logits = logits / self.temperature
# logps = selective_log_softmax(logits, input_ids)
# row_indices, col_indices = torch.where(logps < -20)
# # Method 1: Check if tensors have elements
# if len(row_indices) > 0 and len(col_indices) > 0:
# breakpoint() # Breakpoint triggered here
# print("Found high values!")
# return logps # compute logprobs for the input tokens
def _fix_param_name_to_vllm(self, name, extra_prefixes: Optional[list[str]] = None):
extra_prefixes = extra_prefixes or []
prefixes = ["_checkpoint_wrapped_module."] + extra_prefixes
for prefix in prefixes:
name = name.replace(prefix, "")
return name
def _sync_fsdp1_params_to_vllm(self, module: nn.Module, prefix: str = "", visited=None):
"""Memory-efficient post-order traversal of FSDP modules to extract full parameters and sync with vLLM."""
# For FSDP1, we need to recurse into children and also use summon_full_params
if visited is None:
visited = set()
for child_name, child_module in module.named_children():
child_prefix = f"{prefix}.{child_name}" if prefix else child_name
self._sync_fsdp1_params_to_vllm(
child_module, prefix=child_prefix, visited=visited
) # recurse into the child
if isinstance(module, FSDP):
with FSDP.summon_full_params(module, recurse=False, writeback=False):
for param_name, param in module.named_parameters():
full_name = f"{prefix}.{param_name}" if prefix else param_name
full_name = self._fix_param_name_to_vllm(full_name, extra_prefixes=["_fsdp_wrapped_module."])
if full_name in visited:
continue # skip FSDP subtrees already traversed
visited.add(full_name)
if self.vllm_mode == "server" and self.accelerator.is_main_process:
self.vllm_client.update_named_param(full_name, param.data)
elif self.vllm_mode == "colocate":
pass
pass
def _sync_fsdp2_params_to_vllm(self, module: nn.Module):
# For FSDP2, module already covers all parameters, so no need for recursion
for name, param in module.items():
if param.is_cpu:
param = param.to(torch.device("cuda"))
param = param.full_tensor()
if self.vllm_mode == "server" and self.accelerator.is_main_process:
self.vllm_client.update_named_param(name, param)
elif self.vllm_mode == "colocate":
pass
pass
def _move_model_to_vllm(self, *args, **kwargs):
return None
@profiling_decorator
def _prepare_inputs(
self, generation_batch: dict[str, Union[torch.Tensor, Any]]
) -> dict[str, Union[torch.Tensor, Any]]:
# Prepares inputs for model training/evaluation by managing completion generation and batch handling.
# During training:
# - Receives the local generation batch (Per-GPU batch size × steps per generation)
# from the modified training dataloader instead of the standard local batch
# - Generates completions once for the entire generation batch and splits it into batches of size
# `per_device_train_batch_size`
# - Buffers these completions and returns the appropriate slice for the current accumulation step
# - Optimizes by regenerating completions only periodically (every steps_per_generation * num_iterations)
# During evaluation:
# - The input is treated as a standard local batch (no accumulation, no multiple iterations)
# - Completions are generated for each batch without buffering or reuse
# Returns a single local batch in both cases.
mode = "train" if self.model.training else "eval"
if mode == "train":
generate_every = self.args.steps_per_generation * self.num_iterations
if self._step % generate_every == 0 or self._buffered_inputs is None:
# self._buffered_inputs=None can occur when resuming from a checkpoint
generation_batch = self._generate_and_score_completions(generation_batch)
generation_batch = split_pixel_values_by_grid(generation_batch)
try: generation_batch = shuffle_sequence_dict(generation_batch)
except: pass
generation_batches = split_tensor_dict(generation_batch, self.args.steps_per_generation)
self._buffered_inputs = [unsplit_pixel_values_by_grid(batch) for batch in generation_batches]
inputs = self._buffered_inputs[self._step % self.args.steps_per_generation]
self._step += 1
else:
# In evaluation, there is neither batch grouping for generation, nor multiple iterations, hence
# local generation batch == local eval batch
inputs = self._generate_and_score_completions(generation_batch)
return inputs
@profiling_decorator
def _calculate_rewards(self, inputs, prompts, completions, completion_ids_list):
device = self.accelerator.device
rewards_per_func = torch.zeros(len(prompts), len(self.reward_funcs), device=device)
# Repeat all input columns (but "prompt", "completion", and "completion_ids") to match the num of generations
keys = [key for key in inputs[0] if key not in ["prompt", "completion", "completion_ids"]]
reward_kwargs = {key: [example[key] for example in inputs] for key in keys}
# This allows for dynamic reward shaping based on training progress.
reward_kwargs["trainer_state"] = self.state
for i, (reward_func, reward_processing_class, reward_func_name) in enumerate(
zip(self.reward_funcs, self.reward_processing_classes, self.reward_func_names)
):
with profiling_context(self, reward_func_name):
if isinstance(reward_func, nn.Module): # Module (no PretrainedModel) for compat with compiled models
if is_conversational(inputs[0]):
messages = [{"messages": p + c} for p, c in zip(prompts, completions)]
texts = [apply_chat_template(x, reward_processing_class)["text"] for x in messages]
else:
texts = [p + c for p, c in zip(prompts, completions)]
reward_inputs = reward_processing_class(
text=texts, return_tensors="pt", padding=True, padding_side="right", add_special_tokens=False
)
reward_inputs = super()._prepare_inputs(reward_inputs)
with torch.inference_mode():
rewards_per_func[:, i] = reward_func(**reward_inputs).logits[:, 0] # Shape (B*G,)
else:
output_reward_func = reward_func(
prompts=prompts, completions=completions, completion_ids=completion_ids_list, **reward_kwargs
)
# Convert None values to NaN
output_reward_func = [reward if reward is not None else torch.nan for reward in output_reward_func]
rewards_per_func[:, i] = torch.tensor(output_reward_func, dtype=torch.float32, device=device)
# If all reward functions return None for a given row, issue a detailed warning
if torch.isnan(rewards_per_func).all(dim=1).any():
nan_row_idx = torch.isnan(rewards_per_func).all(dim=1).nonzero(as_tuple=True)[0][0]
row_reward_kwargs = {
key: value[nan_row_idx] for key, value in reward_kwargs.items() if key != "trainer_state"
}
row_reward_kwargs["prompt"] = prompts[nan_row_idx]
row_reward_kwargs["completion"] = completions[nan_row_idx]
logger.warning(
f"All reward functions returned None for the following kwargs:\n{row_reward_kwargs}\n"
"Please ensure that at least one reward function returns a valid reward."
)
# Gather the reward per function: this part is crucial, because the rewards are normalized per group and the
# completions may be distributed across processes
rewards_per_func = gather(rewards_per_func)
return rewards_per_func
def _generate_single_turn(self, prompts: list[str], images: Optional[list]):
device = self.accelerator.device
# If the prompts are conversational and the inputs contain images, we need to convert the prompts from
# [{"role": "user", "content": "What color is the sky?"}] to
# [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": "What color is the sky?"}]}]
kwargs = {}
if images is not None:
kwargs = {"images": images}
for prompt, image_list in zip(prompts, images):
if isinstance(prompt, list): # i.e., when using conversational data
prepare_multimodal_messages(prompt, num_images=len(image_list))
_chat_template_ = getattr(self.processing_class, "chat_template", None)
if _chat_template_ is None: _chat_template_ = ""
_supported_keys_ = set(("prompt", "chosen", "rejected", "completion", "messages", "label"))
_batch_chat_kwargs_ = getattr(self, "_unsloth_batch_chat_kwargs", None)
prompts_text = []
for _idx_, _example_ in enumerate(prompts):
_tokenizer_kwargs_ = {}
if type(_example_) is not dict:
_example_ = {"prompt": _example_}
_left_keys_ = _example_.keys() - _supported_keys_
for k in _left_keys_:
if k in _chat_template_:
v = _example_[k]
if type(v) is str:
_tokenizer_kwargs_[k] = v
if _batch_chat_kwargs_ is not None and _idx_ < len(_batch_chat_kwargs_):
for _bk_, _bv_ in _batch_chat_kwargs_[_idx_].items():
if _bk_ not in _tokenizer_kwargs_:
_tokenizer_kwargs_[_bk_] = _bv_
_x_ = maybe_apply_chat_template(_example_, self.processing_class, **_tokenizer_kwargs_)["prompt"]
prompts_text.append(_x_)
if images is not None:
prompt_inputs = self.processing_class(text=prompts_text, padding=True, return_tensors="pt", **kwargs)
prompt_inputs = super()._prepare_inputs(prompt_inputs)
forward_kwargs = {k: v for k, v in prompt_inputs.items() if k not in ["input_ids", "attention_mask"]}
else:
forward_kwargs = {}
# Generate completions using either vLLM or regular generation
if self.use_vllm:
if self.vllm_mode == "colocate" and self.args.vllm_enable_sleep_mode:
# wake up colocated vLLM instances if needed
torch.cuda.empty_cache() # required to avoid OOM in some cases
self.llm.wake_up()
# First, update the vLLM weights if needed
if self.state.global_step != self._last_loaded_step:
self._move_model_to_vllm()
self._last_loaded_step = self.state.global_step
# Generate completions using vLLM: gather all prompts and use them in a single call in the main process
if self.vllm_mode == "server":
all_prompts_text = gather_object(prompts_text)
if images is not None:
all_images = gather_object(images)
if self.accelerator.is_main_process:
# Since 'prompts' contains 'num_generations' duplicates, we first take unique prompts, and generate
# num_generations outputs for each one. This is faster than generating outputs for each duplicate
# prompt individually.
ordered_set_of_prompts = all_prompts_text[:: self.num_generations]
if images is not None:
ordered_set_of_images = all_images[:: self.num_generations]
else:
ordered_set_of_images = None
with profiling_context(self, "vLLM.generate"):
output = self.vllm_client.generate(
prompts=ordered_set_of_prompts,
images=ordered_set_of_images,
n=self.num_generations,
repetition_penalty=self.repetition_penalty,
temperature=self.temperature,
top_p=self.top_p,
top_k=-1 if self.top_k is None else self.top_k,
min_p=0.0 if self.min_p is None else self.min_p,
max_tokens=self.max_completion_length,
truncate_prompt_tokens=self.max_prompt_length,
guided_decoding_regex=self.guided_decoding_regex,
generation_kwargs=self.args.generation_kwargs,
)
payload = (output["prompt_ids"], output["completion_ids"], output["logprobs"])
else:
payload = None
# Broadcast the completions from the main process to all processes, ensuring each process receives its corresponding slice.
obj_list = [payload]
broadcast_object_list(obj_list, from_process=0)
all_prompt_ids, all_completion_ids, all_logprobs = obj_list[0]
# At this point, we only get 1 copy of each prompt, so we need to repeat them num_generations times
all_prompt_ids = [ids for ids in all_prompt_ids for _ in range(self.num_generations)]
process_slice = slice(
self.accelerator.process_index * len(prompts),
(self.accelerator.process_index + 1) * len(prompts),
)
prompt_ids = all_prompt_ids[process_slice]
completion_ids = all_completion_ids[process_slice]
logprobs = all_logprobs[process_slice]
# Generate completions using colocated vLLM instances: each device holds vLLM copy and work on their own batch of prompts
elif self.vllm_mode == "colocate":
if self.guided_decoding_regex:
guided_decoding = GuidedDecodingParams(regex=self.guided_decoding_regex)
else:
guided_decoding = None
generation_kwargs = {
"n": 1, # vLLM on each GPU generates only 1 in colocate mode
"repetition_penalty": self.repetition_penalty,
"temperature": self.temperature,
"top_p": self.top_p,
"top_k": -1 if self.top_k is None else self.top_k,
"min_p": 0.0 if self.min_p is None else self.min_p,
"max_tokens": self.max_completion_length,
"truncate_prompt_tokens": self.max_prompt_length,
"guided_decoding": guided_decoding,
"logprobs": 0, # only return the logprob of the generated token
}
if self.args.generation_kwargs is not None:
generation_kwargs.update(self.args.generation_kwargs)
sampling_params = SamplingParams(**grpo_update_SamplingParams(SamplingParams, generation_kwargs, getattr(self.args, 'vllm_sampling_params', None)))
if self.vllm_tensor_parallel_size > 1:
# Gather prompts from all ranks in the TP group and flatten.
# Each rank starts with its own prompts; after gathering, all ranks see the full group set.
orig_size = len(prompts_text)
gathered_prompts = [None for _ in range(self.vllm_tensor_parallel_size)]
torch.distributed.all_gather_object(gathered_prompts, prompts_text, group=self.tp_group)
all_prompts_text = [p for sublist in gathered_prompts for p in sublist]
if images is not None:
gathered_images = [None for _ in range(self.vllm_tensor_parallel_size)]
torch.distributed.all_gather_object(gathered_images, images, group=self.tp_group)
all_images = [img for sublist in gathered_images for img in sublist]
else:
all_images = None
else:
all_prompts_text = prompts_text
all_images = images
if images is not None and all_images:
vllm_inputs = []
for prompt, image_list in zip(all_prompts_text, all_images):
vllm_inputs.append({"prompt": prompt, "multi_modal_data": {"image": image_list}})
else:
vllm_inputs = all_prompts_text
with profiling_context(self, "vLLM.generate"):
all_outputs = self.llm.generate(vllm_inputs, sampling_params=sampling_params, use_tqdm=False, lora_request = self.model.load_lora('grpo_trainer_lora_model', load_tensors = True) if getattr(self.llm, 'shared_weights', False) else None)
all_prompt_ids = [output.prompt_token_ids for output in all_outputs]
all_completion_ids = [output.token_ids for outputs in all_outputs for output in outputs.outputs]
all_logprobs = [
[next(iter(lp.values())).logprob for lp in output.logprobs]
for outputs in all_outputs
for output in outputs.outputs
]
if self.vllm_tensor_parallel_size > 1:
# Slice completions for this rank within its TP group.
# Each rank generates all outputs — we keep only our share.
local_rank_in_group = torch.distributed.get_rank(group=self.tp_group)
tp_slice = slice(local_rank_in_group * orig_size, (local_rank_in_group + 1) * orig_size)
prompt_ids = all_prompt_ids[tp_slice]
completion_ids = all_completion_ids[tp_slice]
logprobs = all_logprobs[tp_slice]
else:
prompt_ids = all_prompt_ids
completion_ids = all_completion_ids
logprobs = all_logprobs
if self.args.vllm_enable_sleep_mode:
self.llm.sleep(level=1)
elif self.use_transformers_paged:
# Re-process inputs for paged generation if needed
# Note: images are already validated and preprocessed above
paged_prompt_inputs = self.processing_class(text=prompts_text, **kwargs)
previous_attn = self.model_wrapped.config._attn_implementation
if is_flash_attn_2_available():
self.model_wrapped.config._attn_implementation = "paged_attention"
else:
self.model_wrapped.config._attn_implementation = "sdpa_paged"
with (
profiling_context(self, "transformers.generate_batch"),
unwrap_model_for_generation(
self.model_wrapped, self.accelerator, gather_deepspeed3_params=self.args.ds3_gather_for_generation
) as unwrapped_model,
torch.no_grad(),
FSDP.summon_full_params(self.model_wrapped, recurse=False) if self.is_fsdp_enabled else nullcontext(),
):
# Cast to the appropriate dtype based on training configuration
if self.args.bf16:
unwrapped_model.to(torch.bfloat16)
elif self.args.fp16:
unwrapped_model.to(torch.float16)
with torch.inference_mode():
all_outputs = unwrapped_model.generate_batch(
paged_prompt_inputs.input_ids, generation_config=self.generation_config, progress_bar=False
)
unwrapped_model.train() # restore training mode, as generate_batch forces eval mode
completion_ids = [output.generated_tokens for output in all_outputs.values()]
prompt_ids = paged_prompt_inputs.input_ids
# Restore the original attention implementation, training mode
self.model_wrapped.config._attn_implementation = previous_attn
logprobs = None # not used in this case
else:
# Regular generation path
generate_inputs = self.processing_class(
text=prompts_text,
return_tensors="pt",
padding=True,
padding_side="left",
**kwargs,
)
generate_inputs = super()._prepare_inputs(generate_inputs)
if "mm_token_type_ids" in generate_inputs or "image_grid_thw" in generate_inputs:
mm_token_type_ids = _unsloth_fix_mm_token_type_ids(
self.processing_class,
generate_inputs["input_ids"],
generate_inputs.get("mm_token_type_ids", None),
)
if mm_token_type_ids is not None:
generate_inputs["mm_token_type_ids"] = mm_token_type_ids
with (
profiling_context(self, "transformers.generate"),
unwrap_model_for_generation(
self.model_wrapped, self.accelerator, gather_deepspeed3_params=self.args.ds3_gather_for_generation
) as unwrapped_model,
torch.no_grad(),
FSDP.summon_full_params(self.model_wrapped, recurse=False) if self.is_fsdp_enabled else nullcontext(),
):
prompt_completion_ids = unwrapped_model.generate(
**generate_inputs, generation_config=self.generation_config, disable_compile=True
)
# Compute prompt length and extract completion ids
prompt_ids, prompt_mask = generate_inputs["input_ids"], generate_inputs["attention_mask"]
prompt_length = prompt_ids.size(1)
completion_ids = prompt_completion_ids[:, prompt_length:]
# Mask everything after the first EOS token
is_eos = completion_ids == self.eos_token_id
eos_idx = torch.full((is_eos.size(0),), is_eos.size(1), dtype=torch.long, device=device)
eos_idx[is_eos.any(dim=1)] = is_eos.int().argmax(dim=1)[is_eos.any(dim=1)]
sequence_indices = torch.arange(is_eos.size(1), device=device).expand(is_eos.size(0), -1)
completion_mask = (sequence_indices <= eos_idx.unsqueeze(1)).int()
prompt_ids = [p[m].tolist() for p, m in zip(prompt_ids, prompt_mask.bool())]
completion_ids = [c[m].tolist() for c, m in zip(completion_ids, completion_mask.bool())]
logprobs = None # not used in this case
return prompt_ids, completion_ids, logprobs, forward_kwargs
def _generate(self, prompts: list[str], images: Optional[list]):
device = self.accelerator.device
mode = "train" if self.model.training else "eval"
prompt_ids, completion_ids, logprobs, forward_kwargs = self._generate_single_turn(prompts, images)
# Get completion length per sequence, used for logging
prompt_lengths = torch.tensor([len(ids) for ids in prompt_ids], device=device)
completion_lengths = torch.tensor([len(ids) for ids in completion_ids], device=device)
agg_prompt_lengths = self.accelerator.gather(prompt_lengths)
agg_completion_lengths = self.accelerator.gather(completion_lengths)
total_prompt_tokens = agg_prompt_lengths.sum()
total_completion_tokens = agg_completion_lengths.sum() # = num_items_in_batch, required for the DAPO loss
# Log the metrics
if mode == "train":
self.state.num_input_tokens_seen += (total_prompt_tokens + total_completion_tokens).item()
self._metrics[mode]["num_tokens"] = [self.state.num_input_tokens_seen]
# Log completion lengths, mean, min, max
self._metrics[mode]["completions/mean_length"].append(agg_completion_lengths.float().mean().item())
self._metrics[mode]["completions/min_length"].append(agg_completion_lengths.float().min().item())
self._metrics[mode]["completions/max_length"].append(agg_completion_lengths.float().max().item())
# Identify sequences that terminated with EOS and log their lengths
eos_and_pad = [self.eos_token_id, self.pad_token_id]
is_truncated = torch.tensor([ids[-1] not in eos_and_pad for ids in completion_ids], device=device)
agg_is_truncated = self.accelerator.gather(is_truncated)
self._metrics[mode]["completions/clipped_ratio"].append(agg_is_truncated.float().mean().item())
term_completion_lengths = agg_completion_lengths[~agg_is_truncated]
if len(term_completion_lengths) == 0: # edge case where no terminated sequences are found
term_completion_lengths = torch.zeros(1, device=device)
self._metrics[mode]["completions/mean_terminated_length"].append(term_completion_lengths.float().mean().item())
self._metrics[mode]["completions/min_terminated_length"].append(term_completion_lengths.float().min().item())
self._metrics[mode]["completions/max_terminated_length"].append(term_completion_lengths.float().max().item())
return prompt_ids, completion_ids, total_completion_tokens, logprobs, forward_kwargs
def _generate_and_score_completions(
self, inputs: list[dict[str, Union[torch.Tensor, Any]]]
) -> dict[str, Union[torch.Tensor, Any]]:
device = self.accelerator.device
mode = "train" if self.model.training else "eval"
prompts = [x["prompt"] for x in inputs]
# Unsloth: Extract per-sample chat_template_kwargs before metadata is lost
_ct_ = getattr(self.processing_class, 'chat_template', None) or ''
_sk_ = {'prompt', 'chosen', 'rejected', 'completion', 'messages', 'label',
'images', 'image', 'videos', 'video', 'audios', 'audio'}
self._unsloth_batch_chat_kwargs = []
for _inp_ in inputs:
_kw_ = {}
if isinstance(_inp_, dict):
for _k_ in _inp_.keys() - _sk_:
if _k_ in _ct_ and isinstance(_inp_[_k_], str):
_kw_[_k_] = _inp_[_k_]
self._unsloth_batch_chat_kwargs.append(_kw_)
if "images" in inputs[0]:
images = [example.get("images") for example in inputs]
elif "image" in inputs[0]:
images = [[example.get("image")] if example.get("image") is not None else None for example in inputs]
else:
images = None
# Transformers requires at least one image in the batch, otherwise it throws an error
if images is not None and all(img_list == [] for img_list in images):
images = None
(
prompt_ids_list,
completion_ids_list,
num_items_in_batch,
sampling_per_token_logps_list,
forward_kwargs,
) = self._generate(prompts, images)
# Convert lists of token IDs to padded tensors
prompt_ids = [torch.tensor(ids, device=device) for ids in prompt_ids_list]
prompt_mask = [torch.ones_like(ids, dtype=torch.long) for ids in prompt_ids]
prompt_ids = pad(prompt_ids, padding_value=self.pad_token_id, padding_side="left")
prompt_mask = pad(prompt_mask, padding_value=0, padding_side="left")
completion_ids = [torch.tensor(ids, device=device) for ids in completion_ids_list]
completion_mask = [torch.ones_like(ids, dtype=torch.long) for ids in completion_ids]
completion_ids = pad(completion_ids, padding_value=self.pad_token_id, padding_side="right")
completion_mask = pad(completion_mask, padding_value=0, padding_side="right")
if sampling_per_token_logps_list is not None:
sampling_per_token_logps = [torch.tensor(logps, device=device) for logps in sampling_per_token_logps_list]
sampling_per_token_logps = pad(sampling_per_token_logps, padding_value=0.0, padding_side="right")
else:
sampling_per_token_logps = None
# If mask_truncated_completions is enabled, zero out truncated completions in completion_mask
if self.mask_truncated_completions:
eos_and_pad = [self.eos_token_id, self.pad_token_id]
is_truncated = torch.tensor([ids[-1] not in eos_and_pad for ids in completion_ids_list], device=device)
completion_mask = completion_mask * (~is_truncated).unsqueeze(1).int()
# Concatenate prompt_mask with completion_mask for logit computation
prompt_completion_ids = torch.cat([prompt_ids, completion_ids], dim=1) # (B, P+C)
attention_mask = torch.cat([prompt_mask, completion_mask], dim=1) # (B, P+C)
# If token_type_ids are used, extend them with zeros for the completion part
if "token_type_ids" in forward_kwargs:
token_type_ids = forward_kwargs["token_type_ids"]
forward_kwargs["token_type_ids"] = torch.cat(
[token_type_ids, token_type_ids.new_zeros(completion_ids.shape)], dim=1
)
if "mm_token_type_ids" in forward_kwargs or "image_grid_thw" in forward_kwargs:
_mm_token_type_ids = _unsloth_fix_mm_token_type_ids(
self.processing_class,
prompt_completion_ids,
forward_kwargs.get("mm_token_type_ids", None),
completion_ids = completion_ids,
)
if _mm_token_type_ids is not None:
forward_kwargs["mm_token_type_ids"] = _mm_token_type_ids
logits_to_keep = completion_ids.size(1) # we only need to compute the logits for the completion tokens
max_left_pad = None
batch_size = self.args.per_device_train_batch_size if mode == "train" else self.args.per_device_eval_batch_size
try:
# TRL 0.23.1 and below path
if not has_images:
# Left pad prompt before calculation old and ref hidden states
left_pad_tokens_per_prompt = calculate_pad_tokens_in_prompt(prompt_completion_ids, logits_to_keep, self.processing_class.pad_token_id)
max_left_pad = torch.max(left_pad_tokens_per_prompt).item()
except:
# TRL 0.24.0 and below path
if images is None:
# Left pad prompt before calculation old and ref hidden states
left_pad_tokens_per_prompt = calculate_pad_tokens_in_prompt(prompt_completion_ids, logits_to_keep, self.processing_class.pad_token_id)
max_left_pad = torch.max(left_pad_tokens_per_prompt).item()
self.model.for_training(use_gradient_checkpointing=getattr(self.args, 'gradient_checkpointing', True))
num_images = [len(img_list) for img_list in images] if images is not None else None
with torch.no_grad():
# If the generation and optimization steps are misaligned—i.e., if generation does not occur at the end of
# a full optimizer step (when gradient_accumulation_steps is not a multiple of generate_every)—then the
# samples may come from an earlier version of the model. In that case, we need to track old_per_token_logps
# for importance sampling. If the steps are aligned, importance sampling isn't necessary and we set
# old_per_token_logps to None.
# When using vLLM, we always compute old_per_token_logps for importance sampling, it was shown that the
# distribution mismatch between vLLM and the training model can be large and harm the training.
generate_every = self.args.steps_per_generation * self.num_iterations # generation frequency
if self.args.gradient_accumulation_steps % generate_every != 0 or (
self.use_vllm
):
old_per_token_logps, _ = self._get_per_token_logps_and_entropies(
self.model,
prompt_completion_ids,
attention_mask,
logits_to_keep,
batch_size,
num_images=num_images,
**forward_kwargs, # may contain pixel_values, image_grid_thw, pixel_attention_mask and image_sizes
)
else:
old_per_token_logps = None
# Compute the importance sampling ratio when using vLLM, to correct for potential distribution mismatch
if False and self.use_vllm and self.vllm_importance_sampling_correction:
importance_sampling_ratio = torch.exp(old_per_token_logps - sampling_per_token_logps)
importance_sampling_ratio = torch.clamp(
importance_sampling_ratio, max=self.vllm_importance_sampling_cap
)
# Compute the per-token log probabilities for the reference model
if self.beta != 0.0:
if self.ref_model is not None:
ref_per_token_logps, _ = self._get_per_token_logps_and_entropies(
self.ref_model,
prompt_completion_ids,
attention_mask,
logits_to_keep,
batch_size=batch_size,
num_images=num_images,
**forward_kwargs, # may contain pixel_values, image_grid_thw, pixel_attention_mask and image_sizes
)
else:
with self.accelerator.unwrap_model(self.model).disable_adapter():
ref_per_token_logps, _ = self._get_per_token_logps_and_entropies(
self.model,
prompt_completion_ids,
attention_mask,
logits_to_keep,
batch_size=batch_size,
num_images=num_images,
**forward_kwargs, # may contain pixel_values, image_grid_thw, pixel_attention_mask and image_sizes
)
else:
ref_per_token_logps = None
# Decode
prompts_text = self.processing_class.batch_decode(prompt_ids, skip_special_tokens=True)
completions_text = self.processing_class.batch_decode(completion_ids, skip_special_tokens=True)
if is_conversational(inputs[0]):
completions = []
for prompt, completion in zip(prompts, completions_text):
bootstrap = prompt.pop()["content"] if prompt[-1]["role"] == "assistant" else ""
completions.append([{"role": "assistant", "content": bootstrap + completion}])
else:
completions = completions_text
# Calculate rewards for each reward function. rewards_per_func aggregates rewards across all processes. This is
# important because rewards will be normalized per group, and completions are distributed. We will later slice
# rewards_per_func to extract each process's subset.
if images is not None:
rewards_per_func = self._calculate_rewards(inputs, prompts_text, completions_text, completion_ids_list)
else:
rewards_per_func = self._calculate_rewards(inputs, prompts, completions, completion_ids_list)
# Apply weights to each reward function's output and sum
rewards = (rewards_per_func * self.reward_weights.to(device).unsqueeze(0)).nansum(dim=1)
# Compute grouped-wise rewards
mean_grouped_rewards = rewards.view(-1, self.num_generations).mean(dim=1)
# Normalize the rewards to compute the advantages
mean_grouped_rewards = mean_grouped_rewards.repeat_interleave(self.num_generations, dim=0)
advantages = rewards - mean_grouped_rewards
if self.scale_rewards in ["group", "none"]:
# If self.scale_rewards = "none", we'll still log group level std
std_rewards = rewards.view(-1, self.num_generations).std(dim=1)
std_rewards = std_rewards.repeat_interleave(self.num_generations, dim=0)
elif self.scale_rewards == "batch":
# Compute global std
std_rewards = rewards.std().expand_as(rewards)
else:
raise ValueError(
f"Invalid value for scale_rewards: {self.scale_rewards}. Must be one of 'batch', 'group', or 'none'."
)
is_std_zero = torch.isclose(std_rewards, torch.zeros_like(std_rewards))
if self.scale_rewards != "none":
advantages = advantages / (std_rewards + 1e-4)
# Slice to keep only the local part of the data
process_slice = slice(
self.accelerator.process_index * len(prompts),
(self.accelerator.process_index + 1) * len(prompts),
)
all_process_advantages = advantages.clone() # keep the aggregated advantages for logging
advantages = advantages[process_slice]
# Calculate mean reward per function, but only for samples where the function was applied (non-NaN values)
for i, reward_func_name in enumerate(self.reward_func_names):
mean_rewards = torch.nanmean(rewards_per_func[:, i]).item()
self._metrics[mode][f"rewards/{reward_func_name}/mean"].append(mean_rewards)
std_func_rewards = nanstd(rewards_per_func[:, i]).item()
self._metrics[mode][f"rewards/{reward_func_name}/std"].append(std_func_rewards)
self._metrics[mode]["reward"].append(mean_grouped_rewards.mean().item())
self._metrics[mode]["reward_std"].append(std_rewards.mean().item())
self._metrics[mode]["frac_reward_zero_std"].append(is_std_zero.float().mean().item())
# Log prompt and completion texts
self._logs["prompt"].extend(gather_object(prompts_text))
self._logs["completion"].extend(gather_object(completions_text))
for i, name in enumerate(self.reward_func_names):
self._logs["rewards"][name].extend(rewards_per_func[:, i].tolist())
self._logs["advantages"].extend(all_process_advantages.tolist())
if images is not None:
self._logs["images"].extend(gather_object(images))
if False and self.use_vllm and self.vllm_importance_sampling_correction:
delta = torch.abs(old_per_token_logps - sampling_per_token_logps)
delta = delta[completion_mask.bool()]
mean_delta = torch.mean(delta) if delta.numel() > 0 else torch.tensor(0.0, device=device)
max_delta = torch.max(delta) if delta.numel() > 0 else torch.tensor(0.0, device=device)
self._metrics[mode]["sampling/sampling_logp_difference/mean"].append(
self.accelerator.gather(mean_delta).mean().item()
)
self._metrics[mode]["sampling/sampling_logp_difference/max"].append(
self.accelerator.gather(max_delta).max().item()
)
flat_is_ratio = importance_sampling_ratio[completion_mask.bool()]
min_importance_sampling_ratio = (
torch.min(flat_is_ratio) if flat_is_ratio.numel() > 0 else torch.tensor(0.0, device=device)
)
mean_importance_sampling_ratio = (
torch.mean(flat_is_ratio) if flat_is_ratio.numel() > 0 else torch.tensor(0.0, device=device)
)
max_importance_sampling_ratio = (
torch.max(flat_is_ratio) if flat_is_ratio.numel() > 0 else torch.tensor(0.0, device=device)
)
self._metrics[mode]["sampling/importance_sampling_ratio/min"].append(
nanmin(self.accelerator.gather(min_importance_sampling_ratio)).item()
)
self._metrics[mode]["sampling/importance_sampling_ratio/mean"].append(
self.accelerator.gather(mean_importance_sampling_ratio).nanmean().item()
)
self._metrics[mode]["sampling/importance_sampling_ratio/max"].append(
nanmax(self.accelerator.gather(max_importance_sampling_ratio)).item()
)
output = {
"prompt_ids": prompt_ids,
"prompt_mask": prompt_mask,
"completion_ids": completion_ids,
"completion_mask": completion_mask,
"advantages": advantages,
"num_items_in_batch": num_items_in_batch,
}
if old_per_token_logps is not None:
output["old_per_token_logps"] = old_per_token_logps
if False and self.use_vllm and self.vllm_importance_sampling_correction:
output["importance_sampling_ratio"] = importance_sampling_ratio
if ref_per_token_logps is not None:
output["ref_per_token_logps"] = ref_per_token_logps
if "pixel_values" in forward_kwargs:
output["pixel_values"] = forward_kwargs["pixel_values"]
if "image_grid_thw" in forward_kwargs:
output["image_grid_thw"] = forward_kwargs["image_grid_thw"]
if "pixel_attention_mask" in forward_kwargs:
output["pixel_attention_mask"] = forward_kwargs["pixel_attention_mask"]
if "image_sizes" in forward_kwargs:
output["image_sizes"] = forward_kwargs["image_sizes"]
if "token_type_ids" in forward_kwargs:
output["token_type_ids"] = forward_kwargs["token_type_ids"]
if "mm_token_type_ids" in forward_kwargs:
output["mm_token_type_ids"] = forward_kwargs["mm_token_type_ids"]
if images is not None:
output["num_images"] = num_images
if max_left_pad is not None:
output["max_left_pad"] = torch.tensor(prompt_ids.shape[0] * [max_left_pad]).unsqueeze(-1)
try:
if self.use_vllm and getattr(self, "vllm_importance_sampling_correction", False):
output["sampling_per_token_logps"] = sampling_per_token_logps
except NameError:
output["sampling_per_token_logps"] = None
return output
def compute_liger_loss(self, unwrapped_model, inputs):
# Compute the per-token log probabilities for the model
prompt_ids, prompt_mask = inputs["prompt_ids"], inputs["prompt_mask"]
completion_ids, completion_mask = inputs["completion_ids"], inputs["completion_mask"]
input_ids = torch.cat([prompt_ids, completion_ids], dim=1)
attention_mask = torch.cat([prompt_mask, completion_mask], dim=1)
logits_to_keep = completion_ids.size(1) # we only need to compute the logits for the completion tokens
# Get the last hidden state of the model
last_hidden_state = self._get_last_hidden_state(
unwrapped_model,
input_ids,
attention_mask,
logits_to_keep,
inputs.get("pixel_values"),
inputs.get("image_grid_thw"),
inputs.get("pixel_attention_mask"),
inputs.get("image_sizes"),
)
# compute loss and metrics using liger grpo loss
loss, metrics = self.liger_grpo_loss(
_input=last_hidden_state,
lin_weight=unwrapped_model.lm_head.weight,
selected_token_ids=completion_ids,
attention_mask=completion_mask,
advantages=inputs["advantages"],
bias=unwrapped_model.lm_head.bias,
old_per_token_logps=inputs.get("old_per_token_logps"),
ref_per_token_logps=inputs.get("ref_per_token_logps"),
)
# Extract metrics from the liger_grpo_loss output
# KL divergence is the first metric when beta is non-zero
mean_kl = metrics[0] if self.beta != 0.0 else None
clip_ratio = metrics[-1]
mode = "train" if self.model.training else "eval"
if self.beta != 0.0:
self._metrics[mode]["kl"].append(self.accelerator.gather(mean_kl).mean().item())
self._metrics[mode]["clip_ratio"].append(self.accelerator.gather(clip_ratio).mean().item())
return loss / self.current_gradient_accumulation_steps
def compute_loss(
self, model, inputs, return_outputs = False, num_items_in_batch = None
):
if return_outputs:
raise ValueError("The GRPOTrainer does not support returning outputs")
# Compute the per-token log probabilities for the model
prompt_ids, prompt_mask = inputs["prompt_ids"], inputs["prompt_mask"]
completion_ids, completion_mask = (
inputs["completion_ids"],
inputs["completion_mask"],
)
pixel_values, image_grid_thw = (
inputs.get("pixel_values", None),
inputs.get("image_grid_thw", None),
)
pixel_attention_mask, image_sizes = (
inputs.get("pixel_attention_mask", None),
inputs.get("image_sizes", None),
)
num_images = inputs.get("num_images", None)
# Transformers 5.x needs token_type_ids/mm_token_type_ids for some vision models
token_type_ids = inputs.get("token_type_ids", None)
mm_token_type_ids = inputs.get("mm_token_type_ids", None)
num_items_in_batch = inputs.get("num_items_in_batch", None)
sampling_per_token_logps = inputs.get("sampling_per_token_logps", None)
tool_mask = inputs.get("tool_mask", None)
current_gradient_accumulation_steps = self.current_gradient_accumulation_steps
num_processes = self.accelerator.num_processes
input_ids = torch.cat([prompt_ids, completion_ids], dim = 1)
bsz, qlen = input_ids.shape
attention_mask = torch.cat([prompt_mask, completion_mask], dim = 1)
if mm_token_type_ids is not None or image_grid_thw is not None:
mm_token_type_ids = _unsloth_fix_mm_token_type_ids(
self.processing_class,
input_ids,
mm_token_type_ids,
completion_ids = completion_ids,
)
# attention_mask = None
logits_to_keep = completion_ids.size(
1
) # we only need to compute the logits for the completion tokens
_input_ids = input_ids
_logits_to_keep = logits_to_keep
get_logps_func = (
lambda model,
input_ids,
attention_mask,
logits_to_keep,
batch_size = None,
compute_entropy = False,
compute_efficient = False: self._get_per_token_logps(
model, input_ids, attention_mask, logits_to_keep, compute_efficient
)
if hasattr(self, "_get_per_token_logps")
else self._get_per_token_logps_and_entropies(
model,
input_ids,
attention_mask,
logits_to_keep,
batch_size,
compute_entropy,
compute_efficient,
)[0]
) # logps
per_token_logps = get_logps_func(
model, input_ids, attention_mask, logits_to_keep, compute_efficient = True
)
# Compute the KL divergence between the model and the reference model
# _prepare_inputs doesn't return reference log probs anymore. We need to calculate it ourselves.
# https://github.com/huggingface/trl/blob/05bc43e960396581e458195b8388efe6b82cae1f/trl/trainer/grpo_trainer.py#L1328
# if self.beta != 0.0:
# with torch.inference_mode(), model.disable_adapter():
# ref_per_token_logps = per_token_logps = get_logps_func(model, input_ids, attention_mask, logits_to_keep)
# else:
# ref_per_token_logps = None
ref_logps = inputs.get("ref_per_token_logps", None)
# per_token_kl = torch.exp(ref_per_token_logps - per_token_logps) - (ref_per_token_logps - per_token_logps) - 1
# x - x.detach() allows for preserving gradients from x
advantages = inputs["advantages"]
# per_token_loss = torch.exp(per_token_logps - per_token_logps.detach()) * advantages.unsqueeze(1)
# per_token_loss = -(per_token_loss - self.beta * per_token_kl)
# loss = ((per_token_loss * completion_mask).sum(dim=1) / completion_mask.sum(dim=1)).mean()
old_logps = inputs.get("old_per_token_logps", None)
input_ids = input_ids[:, -logits_to_keep:]
# Get logit softcapping and logit scale
logit_softcapping = _unsloth_get_final_logit_softcapping(model.config) # Gemma
logit_scale_multiply = getattr(model.config, "logit_scale", 0) # Cohere
if logit_scale_multiply is None:
logit_scale_multiply = 0
logit_scale_divide = getattr(model.config, "logits_scaling", 0) # Granite
if logit_scale_divide is None:
logit_scale_divide = 0
max_left_pad = inputs.get("max_left_pad", 0)
if per_token_logps is not None:
loss_mask = completion_mask
if tool_mask is not None:
if tool_mask.shape != completion_mask.shape:
raise ValueError(
"tool_mask/env_mask must have the same shape as completion_mask"
)
loss_mask = completion_mask * tool_mask.to(
device = completion_mask.device,
dtype = completion_mask.dtype,
)
(
loss,
completion_length,
mean_kl,
delta,
flat_is_ratio,
coef_1,
completion_mask,
) = grpo_compute_loss_slow(
ref_logps,
per_token_logps,
old_logps,
sampling_per_token_logps,
input_ids,
loss_mask,
self.beta,
advantages,
pixel_values = pixel_values,
image_grid_thw = image_grid_thw,
loss_type = self.args.loss_type,
importance_sampling_level = self.importance_sampling_level,
epsilon_low = self.epsilon_low,
epsilon_high = self.epsilon_high,
max_completion_length = self.args.max_completion_length,
delta = self.args.delta,
temperature = self.args.temperature,
max_left_pad = max_left_pad,
logit_softcapping = logit_softcapping,
logit_scale_multiply = logit_scale_multiply,
logit_scale_divide = logit_scale_divide,
num_items_in_batch = num_items_in_batch,
current_gradient_accumulation_steps = current_gradient_accumulation_steps,
num_processes = num_processes,
)
else:
def _unsloth_requires_multi_image_zoo(value):
if value is None:
return False
if isinstance(value, torch.Tensor):
counts = value.detach().cpu().reshape(-1).tolist()
else:
counts = list(value)
return any(int(n) != 1 for n in counts)
if _unsloth_requires_multi_image_zoo(num_images) and not getattr(
self, "_unsloth_grpo_zoo_checked", False
):
_supports_num_images = (
"num_images" in inspect.signature(grpo_accumulated_loss).parameters
)
if not _supports_num_images:
try:
_zoo_src = inspect.getsource(grpo_accumulated_loss)
except (TypeError, OSError):
_zoo_src = ""
_supports_num_images = "num_images" in _zoo_src
if not _supports_num_images:
raise RuntimeError(
"Multi-image GRPO requires an unsloth_zoo build whose "
"grpo_accumulated_loss handles num_images. Please upgrade "
"unsloth_zoo (see https://github.com/unslothai/unsloth-zoo/pull/613)."
)
self._unsloth_grpo_zoo_checked = True
if tool_mask is not None and not getattr(
self, "_unsloth_grpo_tool_mask_zoo_checked", False
):
_supports_tool_mask = (
"tool_mask" in inspect.signature(grpo_accumulated_loss).parameters
)
if not _supports_tool_mask:
try:
_zoo_src = inspect.getsource(grpo_accumulated_loss)
except (TypeError, OSError):
_zoo_src = ""
_supports_tool_mask = "tool_mask" in _zoo_src
if not _supports_tool_mask:
raise RuntimeError(
"env_mask/tool_mask GRPO requires an unsloth_zoo build whose "
"grpo_accumulated_loss handles tool_mask. Please upgrade "
"unsloth_zoo."
)
self._unsloth_grpo_tool_mask_zoo_checked = True
_grpo_accumulated_loss_kwargs = {}
if tool_mask is not None:
_grpo_accumulated_loss_kwargs["tool_mask"] = tool_mask
if hasattr(self.args, "loss_type"):
(
loss,
completion_length,
mean_kl,
delta,
flat_is_ratio,
coef_1,
completion_mask,
) = grpo_accumulated_loss(
trainer = self,
input_ids = _input_ids,
pixel_values = pixel_values,
image_grid_thw = image_grid_thw,
pixel_attention_mask = pixel_attention_mask,
image_sizes = image_sizes,
num_images = num_images,
logits_to_keep = logits_to_keep,
completion_mask = completion_mask,
advantages = advantages,
old_logps = old_logps,
ref_logps = ref_logps,
n_chunks = self.args.unsloth_num_chunks,
loss_type = self.args.loss_type,
importance_sampling_level = self.importance_sampling_level,
epsilon_low = self.epsilon_low,
epsilon_high = self.epsilon_high,
max_completion_length = self.args.max_completion_length,
delta = self.args.delta,
temperature = self.args.temperature,
max_left_pad = max_left_pad,
logit_softcapping = logit_softcapping,
logit_scale_multiply = logit_scale_multiply,
logit_scale_divide = logit_scale_divide,
attention_mask = attention_mask,
num_items_in_batch = num_items_in_batch,
current_gradient_accumulation_steps = current_gradient_accumulation_steps,
num_processes = num_processes,
sampling_per_token_logps = sampling_per_token_logps,
token_type_ids = token_type_ids,
mm_token_type_ids = mm_token_type_ids,
**_grpo_accumulated_loss_kwargs,
)
else:
# to ensure backwards compatibility with trl 0.15.2 and maybe even 0.17
loss, completion_length, mean_kl, coef_1, completion_mask = (
grpo_accumulated_loss(
trainer = self,
input_ids = _input_ids,
pixel_values = pixel_values,
image_grid_thw = image_grid_thw,
pixel_attention_mask = pixel_attention_mask,
image_sizes = image_sizes,
num_images = num_images,
logits_to_keep = logits_to_keep,
completion_mask = completion_mask,
advantages = advantages,
old_logps = old_logps,
ref_logps = ref_logps,
n_chunks = self.args.unsloth_num_chunks,
temperature = self.args.temperature,
logit_softcapping = logit_softcapping,
logit_scale_multiply = logit_scale_multiply,
logit_scale_divide = logit_scale_divide,
attention_mask = attention_mask,
token_type_ids = token_type_ids,
mm_token_type_ids = mm_token_type_ids,
**_grpo_accumulated_loss_kwargs,
)
)
if "train" in self._metrics:
mode = "eval" if self.control.should_evaluate else "train"
self._metrics[mode]["completion_length"].append(completion_length.item())
self._metrics[mode]["kl"].append(mean_kl.item())
else:
self._metrics["completion_length"].append(completion_length.item())
self._metrics["kl"].append(mean_kl.item())
if (
self.use_vllm
and delta is not None
and getattr(self, "vllm_importance_sampling_correction", False)
):
mean_delta = (
torch.mean(delta)
if delta.numel() > 0
else torch.tensor(0.0, device = self.model.device)
)
max_delta = (
torch.max(delta)
if delta.numel() > 0
else torch.tensor(0.0, device = self.model.device)
)
self._metrics[mode]["sampling/sampling_logp_difference/mean"].append(
self.accelerator.gather(mean_delta).mean().item()
)
self._metrics[mode]["sampling/sampling_logp_difference/max"].append(
self.accelerator.gather(max_delta).max().item()
)
min_importance_sampling_ratio = (
torch.min(flat_is_ratio)
if flat_is_ratio.numel() > 0
else torch.tensor(0.0, device = self.model.device)
)
mean_importance_sampling_ratio = (
torch.mean(flat_is_ratio)
if flat_is_ratio.numel() > 0
else torch.tensor(0.0, device = self.model.device)
)
max_importance_sampling_ratio = (
torch.max(flat_is_ratio)
if flat_is_ratio.numel() > 0
else torch.tensor(0.0, device = self.model.device)
)
self._metrics[mode]["sampling/importance_sampling_ratio/min"].append(
self.accelerator.gather(min_importance_sampling_ratio)
.nan_to_num(nan = float("inf"))
.min()
.item()
)
self._metrics[mode]["sampling/importance_sampling_ratio/mean"].append(
self.accelerator.gather(mean_importance_sampling_ratio).nanmean().item()
)
self._metrics[mode]["sampling/importance_sampling_ratio/max"].append(
self.accelerator.gather(max_importance_sampling_ratio)
.nan_to_num(nan = float("-inf"))
.max()
.item()
)
completion_token_count = completion_mask.sum().clamp(min = 1.0)
def masked_batch_mean(x):
if x.shape[1] == 1: # when importance_sampling_level == "sequence"
return x.mean()
else:
return (x * completion_mask).sum() / completion_token_count
if advantages.dim() == 1:
advantages = advantages.unsqueeze(1)
if self.loss_type in ["grpo", "bnpo", "dr_grpo", "dapo"]:
# Compute the clipped probability ratios
is_low_clipped = (coef_1 < 1 - self.epsilon_low) & (advantages < 0)
is_high_clipped = (coef_1 > 1 + self.epsilon_high) & (advantages > 0)
is_region_clipped = is_low_clipped | is_high_clipped
low_clip = masked_batch_mean(is_low_clipped.float())
high_clip = masked_batch_mean(is_high_clipped.float())
clip_ratio = masked_batch_mean(is_region_clipped.float())
gathered_low_clip = self.accelerator.gather(low_clip)
self._metrics[mode]["clip_ratio/low_mean"].append(
gathered_low_clip.nanmean().item()
)
self._metrics[mode]["clip_ratio/low_min"].append(
nanmin(gathered_low_clip).item()
)
gathered_high_clip = self.accelerator.gather(high_clip)
self._metrics[mode]["clip_ratio/high_mean"].append(
gathered_high_clip.nanmean().item()
)
self._metrics[mode]["clip_ratio/high_max"].append(
nanmax(gathered_high_clip).item()
)
gathered_clip_ratio = self.accelerator.gather(clip_ratio)
self._metrics[mode]["clip_ratio/region_mean"].append(
gathered_clip_ratio.nanmean().item()
)
elif self.loss_type == "cispo":
is_cispo_clipped = (coef_1 > self.epsilon_high) & (advantages > 0)
cispo_clip_ratio = masked_batch_mean(is_cispo_clipped.float())
gathered_cispo_clip_ratio = self.accelerator.gather(cispo_clip_ratio)
self._metrics[mode]["cispo_clip_ratio"].append(
gathered_cispo_clip_ratio.nanmean().item()
)
return loss
def _compute_loss(self, model, inputs):
# Compute the per-token log probabilities for the model
prompt_ids, prompt_mask = inputs["prompt_ids"], inputs["prompt_mask"]
completion_ids, completion_mask = inputs["completion_ids"], inputs["completion_mask"]
input_ids = torch.cat([prompt_ids, completion_ids], dim=1)
attention_mask = torch.cat([prompt_mask, completion_mask], dim=1)
logits_to_keep = completion_ids.size(1) # we only need to compute the logits for the completion tokens
# Compute the per_token_logps and the entropy at each position in the completion
per_token_logps, entropies = self._get_per_token_logps_and_entropies(
model,
input_ids,
attention_mask,
logits_to_keep,
compute_entropy=True,
pixel_values=inputs.get("pixel_values"),
image_grid_thw=inputs.get("image_grid_thw"),
num_images=inputs.get("num_images"),
pixel_attention_mask=inputs.get("pixel_attention_mask"),
image_sizes=inputs.get("image_sizes"),
token_type_ids=inputs.get("token_type_ids"),
)
if self.top_entropy_quantile < 1.0:
entropy_mask = self.get_high_entropy_mask(entropies, completion_mask, 1 - self.top_entropy_quantile)
else:
entropy_mask = None
# Compute the KL divergence between the model and the reference model
if self.beta != 0.0:
ref_per_token_logps = inputs["ref_per_token_logps"]
per_token_kl = (
torch.exp(ref_per_token_logps - per_token_logps) - (ref_per_token_logps - per_token_logps) - 1
)
# Compute the loss
advantages = inputs["advantages"]
# When num_iterations == 1 and steps_per_generation <= gradient_accumulation_steps,
# old_per_token_logps == per_token_logps. In this case we can skip its computation
# (see _generate_and_score_completions) and instead use per_token_logps.detach().
# The exception is when using vLLM, where we always compute old_per_token_logps
# for importance sampling
old_per_token_logps = inputs.get("old_per_token_logps")
old_per_token_logps = per_token_logps.detach() if old_per_token_logps is None else old_per_token_logps
log_ratio = per_token_logps - old_per_token_logps
if self.importance_sampling_level == "token":
log_importance_weights = log_ratio
elif self.importance_sampling_level == "sequence":
log_importance_weights = (log_ratio * completion_mask).sum(-1) / completion_mask.sum(-1).clamp(min=1.0)
log_importance_weights = log_importance_weights.unsqueeze(-1)
else:
raise ValueError(
f"Unknown importance sampling level: {self.importance_sampling_level}. Possible values are 'token' "
"and 'sequence'."
)
# From here, log_importance_weights (and all subsequent tensors, coef_1, coef_2, etc.) shape depends on
# importance_sampling_level: "token" level: (B, T); "sequence" level: (B, 1)
coef_1 = torch.exp(log_importance_weights)
coef_2 = torch.clamp(coef_1, 1 - self.epsilon_low, 1 + self.epsilon_high)
# Two-sided clipping
if self.args.delta is not None:
coef_1 = torch.clamp(coef_1, max=self.args.delta)
per_token_loss1 = coef_1 * advantages.unsqueeze(1)
per_token_loss2 = coef_2 * advantages.unsqueeze(1)
per_token_loss = -torch.min(per_token_loss1, per_token_loss2)
if entropy_mask is not None:
per_token_loss = per_token_loss * entropy_mask
if self.use_vllm and self.vllm_importance_sampling_correction:
per_token_loss = per_token_loss * inputs["importance_sampling_ratio"]
if self.beta != 0.0:
per_token_loss = per_token_loss + self.beta * per_token_kl
if self.loss_type == "grpo":
loss = ((per_token_loss * completion_mask).sum(-1) / completion_mask.sum(-1).clamp(min=1.0)).mean()
loss = loss / self.current_gradient_accumulation_steps
elif self.loss_type == "bnpo":
loss = (per_token_loss * completion_mask).sum() / completion_mask.sum().clamp(min=1.0)
loss = loss / self.current_gradient_accumulation_steps
elif self.loss_type == "dr_grpo":
loss = (per_token_loss * completion_mask).sum() / (per_token_loss.size(0) * self.max_completion_length)
loss = loss / self.current_gradient_accumulation_steps
elif self.loss_type == "dapo":
normalizer = inputs["num_items_in_batch"] / self.accelerator.num_processes
loss = (per_token_loss * completion_mask).sum() / normalizer
else:
raise ValueError(f"Unknown loss type: {self.loss_type}")
# Log the metrics
mode = "train" if self.model.training else "eval"
completion_token_count = completion_mask.sum().clamp(min=1.0)
def masked_batch_mean(x):
if x.shape[1] == 1: # when importance_sampling_level == "sequence"
return x.mean()
else:
return (x * completion_mask).sum() / completion_token_count
if self.beta != 0.0:
mean_kl = masked_batch_mean(per_token_kl)
self._metrics[mode]["kl"].append(self.accelerator.gather(mean_kl).nanmean().item())
mean_entropy = masked_batch_mean(entropies)
self._metrics[mode]["entropy"].append(self.accelerator.gather(mean_entropy).nanmean().item())
# Compute the clipped probability ratios
is_low_clipped = (coef_1 < 1 - self.epsilon_low) & (advantages.unsqueeze(1) < 0)
is_high_clipped = (coef_1 > 1 + self.epsilon_high) & (advantages.unsqueeze(1) > 0)
is_region_clipped = is_low_clipped | is_high_clipped
low_clip = masked_batch_mean(is_low_clipped.float())
high_clip = masked_batch_mean(is_high_clipped.float())
clip_ratio = masked_batch_mean(is_region_clipped.float())
gathered_low_clip = self.accelerator.gather(low_clip)
self._metrics[mode]["clip_ratio/low_mean"].append(gathered_low_clip.nanmean().item())
self._metrics[mode]["clip_ratio/low_min"].append(nanmin(gathered_low_clip).item())
gathered_high_clip = self.accelerator.gather(high_clip)
self._metrics[mode]["clip_ratio/high_mean"].append(gathered_high_clip.nanmean().item())
self._metrics[mode]["clip_ratio/high_max"].append(nanmax(gathered_high_clip).item())
gathered_clip_ratio = self.accelerator.gather(clip_ratio)
self._metrics[mode]["clip_ratio/region_mean"].append(gathered_clip_ratio.nanmean().item())
return loss
def prediction_step(self, model, inputs, prediction_loss_only, ignore_keys: Optional[list[str]] = None):
inputs = self._prepare_inputs(inputs)
with torch.no_grad():
with self.compute_loss_context_manager():
loss = self.compute_loss(model, inputs)
loss = loss.mean().detach()
return loss, None, None
def log(self, logs: dict[str, float], start_time: Optional[float] = None) -> None:
mode = "train" if self.model.training else "eval"
metrics = {key: sum(val) / len(val) for key, val in self._metrics[mode].items()} # average the metrics
# This method can be called both in training and evaluation. When called in evaluation, the keys in `logs`
# start with "eval_". We need to add the prefix "eval_" to the keys in `metrics` to match the format.
if mode == "eval":
metrics = {f"eval_{key}": val for key, val in metrics.items()}
logs = {**logs, **metrics}
super().log(logs, start_time)
self._metrics[mode].clear()
if self.accelerator.is_main_process and self.log_completions:
if is_rich_available():
print_prompt_completions_sample(
self._logs["prompt"],
self._logs["completion"],
self._logs["rewards"],
self._logs["advantages"],
self.state.global_step,
self.num_completions_to_print,
)
if self.args.report_to and "wandb" in self.args.report_to and wandb.run is not None:
import pandas as pd
table = {
"step": [str(self.state.global_step)] * len(self._logs["prompt"]),
"prompt": self._logs["prompt"],
"completion": self._logs["completion"],
**self._logs["rewards"],
"advantage": self._logs["advantages"],
}
if self._logs["images"]:
table["images"] = []
for image_list in self._logs["images"]:
# Convert images to wandb Image objects for proper visualization
table["images"].append([wandb.Image(image) for image in image_list])
df = pd.DataFrame(table)
if self.wandb_log_unique_prompts:
df = df.drop_duplicates(subset=["prompt"])
wandb.log({"completions": wandb.Table(dataframe=df)})
# Ensure the model card is saved along with the checkpoint
def _save_checkpoint(self, model, trial):
if self.args.hub_model_id is None:
model_name = Path(self.args.output_dir).name
else:
model_name = self.args.hub_model_id.split("/")[-1]
self.create_model_card(model_name=model_name)
super()._save_checkpoint(model, trial)
class UnslothGRPOTrainer(_UnslothGRPOTrainer):
"""
Trainer for the Group Relative Policy Optimization (GRPO) method. This algorithm was initially proposed in the
paper [DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language
Models](https://huggingface.co/papers/2402.03300).
Example:
```python
from datasets import load_dataset
from trl import GRPOTrainer
dataset = load_dataset("trl-lib/tldr", split="train")
def reward_func(completions, **kwargs):
# Dummy reward function that rewards completions with more unique letters.
return [float(len(set(completion))) for completion in completions]
trainer = GRPOTrainer(
model="Qwen/Qwen2-0.5B-Instruct",
reward_funcs=reward_func,
train_dataset=dataset,
)
trainer.train()
```
Args:
model (`Union[str, PreTrainedModel]`):
Model to be trained. Can be either:
- A string, being the *model id* of a pretrained model hosted inside a model repo on huggingface.co, or a
path to a *directory* containing model weights saved using
[`~transformers.PreTrainedModel.save_pretrained`], e.g., `'./my_model_directory/'`. The model is loaded
using [`~transformers.AutoModelForCausalLM.from_pretrained`] with the keyword arguments in
`args.model_init_kwargs`.
- A [`~transformers.PreTrainedModel`] object. Only causal language models are supported.
reward_funcs (`Union[RewardFunc, list[RewardFunc]]`):
Reward functions to be used for computing the rewards. To compute the rewards, we call all the reward
functions with the prompts and completions and sum the rewards. Can be either:
- A single reward function, such as:
- A string: The *model ID* of a pretrained model hosted inside a model repo on huggingface.co, or a
path to a *directory* containing model weights saved using
[`~transformers.PreTrainedModel.save_pretrained`], e.g., `'./my_model_directory/'`. The model is loaded
using [`~transformers.AutoModelForSequenceClassification.from_pretrained`] with `num_labels=1` and the
keyword arguments in `args.model_init_kwargs`.
- A [`~transformers.PreTrainedModel`] object: Only sequence classification models are supported.
- A custom reward function: The function is provided with the prompts and the generated completions,
plus any additional columns in the dataset. It should return a list of rewards. Custom reward
functions can also return `None` when the reward is not applicable to those samples. This is useful
for multi-task training where different reward functions apply to different types of samples. When a
reward function returns `None` for a sample, that reward function is excluded from the reward
calculation for that sample. For more details, see [Using a custom reward
function](#using-a-custom-reward-function).
The trainer's state is also passed to the reward function. The trainer's state is an instance of
[`~transformers.TrainerState`] and can be accessed by accessing the `trainer_state` argument to the
reward function's signature.
- A list of reward functions, where each item can independently be any of the above types. Mixing different
types within the list (e.g., a string model ID and a custom reward function) is allowed.
args ([`GRPOConfig`], *optional*):
Configuration for this trainer. If `None`, a default configuration is used.
train_dataset ([`~datasets.Dataset`] or [`~datasets.IterableDataset`]):
Dataset to use for training. It must include a column `"prompt"`. Any additional columns in the dataset is
ignored. The format of the samples can be either:
- [Standard](dataset_formats#standard): Each sample contains plain text.
- [Conversational](dataset_formats#conversational): Each sample contains structured messages (e.g., role
and content).
eval_dataset ([`~datasets.Dataset`], [`~datasets.IterableDataset`] or `dict[str, Union[Dataset, IterableDataset]]`):
Dataset to use for evaluation. It must meet the same requirements as `train_dataset`.
processing_class ([`~transformers.PreTrainedTokenizerBase`], [`~transformers.ProcessorMixin`], *optional*):
Processing class used to process the data. The padding side must be set to "left". If `None`, the
processing class is loaded from the model's name with [`~transformers.AutoProcessor.from_pretrained`]. A
padding token, `tokenizer.pad_token`, must be set. If the processing class has not set a padding token,
`tokenizer.eos_token` will be used as the default.
reward_processing_classes ([`~transformers.PreTrainedTokenizerBase`] or `list[PreTrainedTokenizerBase]`, *optional*):
Processing classes corresponding to the reward functions specified in `reward_funcs`. Can be either:
- A single processing class: Used when `reward_funcs` contains only one reward function.
- A list of processing classes: Must match the order and length of the reward functions in `reward_funcs`.
If set to `None`, or if an element of the list corresponding to a [`~transformers.PreTrainedModel`] is
`None`, the tokenizer for the model is automatically loaded using
[`~transformers.AutoTokenizer.from_pretrained`]. For elements in `reward_funcs` that are custom reward
functions (not [`~transformers.PreTrainedModel`]), the corresponding entries in `reward_processing_classes`
are ignored.
callbacks (list of [`~transformers.TrainerCallback`], *optional*):
List of callbacks to customize the training loop. Will add those to the list of default callbacks detailed
in [here](https://huggingface.co/docs/transformers/main_classes/callback).
If you want to remove one of the default callbacks used, use the [`~transformers.Trainer.remove_callback`]
method.
optimizers (`tuple[torch.optim.Optimizer, torch.optim.lr_scheduler.LambdaLR]`, *optional*, defaults to `(None, None)`):
A tuple containing the optimizer and the scheduler to use. Will default to an instance of [`AdamW`] on your
model and a scheduler given by [`get_linear_schedule_with_warmup`] controlled by `args`.
peft_config ([`~peft.PeftConfig`], *optional*):
PEFT configuration used to wrap the model. If `None`, the model is not wrapped.
"""
def __init__(
self,
model,
reward_funcs,
args = None,
train_dataset = None,
eval_dataset = None,
processing_class = None,
reward_processing_classes = None,
callbacks = None,
peft_config = None,
**kwargs
):
if args is None: args = UnslothGRPOConfig()
use_bf16 = getattr(args, 'bf16', False)
if type(use_bf16) is not bool: use_bf16 = False
use_fp16 = getattr(args, 'fp16', False)
if type(use_fp16) is not bool: use_fp16 = False
force_float32 = False
full_finetuning = os.environ.get('UNSLOTH_ENABLE_FULL_FINETUNING', '0') == '1'
if not full_finetuning and (os.environ.get('UNSLOTH_FORCE_FLOAT32', '0') == '1'):
print('Unsloth: Switching to float32 training since model cannot work with float16')
force_float32 = True
mixed_precision_dtype = os.environ.get('UNSLOTH_MIXED_PRECISION', 'float32')
dtype = getattr(model.config, 'dtype', None) or getattr(model.config, 'torch_dtype', None)
if dtype is None: dtype = model.get_input_embeddings().weight.dtype
from unsloth_zoo.utils import _get_dtype
dtype = _get_dtype(dtype)
float16 = dtype == torch.float16
if not force_float32 and (float16 and use_bf16): raise TypeError('Unsloth: Model is in float16 precision but you want to use bfloat16 precision. Set fp16 to `True` and bf16 to `False`')
if not force_float32 and (not float16 and use_fp16): raise TypeError('Unsloth: Model is in bfloat16 precision but you want to use float16 precision. Set fp16 to `False` and bf16 to `True`')
if force_float32:
# Forced float32 training
args.fp16 = False
args.bf16 = False
os.environ['ACCELERATE_MIXED_PRECISION'] = 'no'
if hasattr(args, 'mixed_precision'): args.mixed_precision = 'no'
# args.mixed_precision is a new argument which needs to be set now
elif (not use_bf16 and not use_fp16) and mixed_precision_dtype == 'float32':
# Mixed precision training
args.fp16 = float16
args.bf16 = not float16
os.environ['ACCELERATE_MIXED_PRECISION'] = 'fp16' if float16 else 'bf16'
if hasattr(args, 'mixed_precision'): args.mixed_precision = 'fp16' if float16 else 'bf16'
# args.mixed_precision is a new argument which needs to be set now
elif mixed_precision_dtype == 'bfloat16':
# Both False since bfloat16 full finetuning doesn't do any autocasting.
args.fp16 = False
args.bf16 = False
os.environ['ACCELERATE_MIXED_PRECISION'] = 'no'
if hasattr(args, 'mixed_precision'): args.mixed_precision = 'no'
# args.mixed_precision is a new argument which needs to be set now
if getattr(args, 'eval_dataset', None) is not None and getattr(args, 'eval_strategy', 'no') == 'no':
args.eval_strategy = 'steps'
if getattr(args, 'eval_steps', None) is None: args.eval_steps = 0.1
ga_steps = getattr(args, 'gradient_accumulation_steps', None)
if ga_steps is not None and ga_steps > 1:
from transformers import __version__ as transformers_version
if Version(transformers_version) <= Version('4.45.2'):
print('**** Unsloth: Please use our fixed gradient_accumulation_steps by updating transformers, TRL and Unsloth!\n'
'`pip install --upgrade --no-cache-dir --force-reinstall --no-deps unsloth transformers trl unsloth_zoo`')
if getattr(args, 'eval_strategy', 'no') != 'no':
eval_bsz = getattr(args, 'per_device_eval_batch_size', 8)
if eval_bsz == 8 and args.per_device_train_batch_size < eval_bsz: args.per_device_eval_batch_size = args.per_device_train_batch_size
if getattr(args, 'eval_accumulation_steps', None) is None and ga_steps is not None: args.eval_accumulation_steps = ga_steps
fp16_full_eval = getattr(args, 'fp16_full_eval', False)
if type(fp16_full_eval) is not bool: fp16_full_eval = False
bf16_full_eval = getattr(args, 'bf16_full_eval', False)
if type(bf16_full_eval) is not bool: bf16_full_eval = False
if args.fp16 and bf16_full_eval: args.bf16_full_eval = False; args.fp16_full_eval = True
if args.bf16 and fp16_full_eval: args.bf16_full_eval = True; args.fp16_full_eval = False
if force_float32:
args.bf16_full_eval = False
args.fp16_full_eval = False
elif os.environ.get('UNSLOTH_MIXED_PRECISION', 'float32') == 'bfloat16':
args.bf16_full_eval = True
args.fp16_full_eval = False
elif not bf16_full_eval and not fp16_full_eval:
args.bf16_full_eval = args.bf16
args.fp16_full_eval = args.fp16
_output_logits = False
if locals().get('compute_metrics', None) is not None: _output_logits = True
if locals().get('preprocess_logits_for_metrics', None) is not None: _output_logits = True
if _output_logits:
os.environ['UNSLOTH_RETURN_LOGITS'] = '1'
if model is not None:
_warnings_issued = getattr(model, 'warnings_issued', None)
if _warnings_issued is None:
model.warnings_issued = {}
elif not isinstance(_warnings_issued, dict):
try:
model.warnings_issued = dict(_warnings_issued)
except Exception:
model.warnings_issued = {}
if 'max_seq_length' not in locals() and not hasattr(args, 'max_seq_length'):
pass
else:
model_max_seq_length = getattr(model, 'max_seq_length', None)
args_max_seq_length = getattr(args, 'max_seq_length', None)
if args_max_seq_length is None and model_max_seq_length is not None:
max_seq_length = model.max_seq_length
if hasattr(args, 'max_seq_length'): args.max_seq_length = max_seq_length
elif args_max_seq_length is not None and model_max_seq_length is not None:
if args_max_seq_length > model_max_seq_length:
print('Unsloth: You set `max_seq_length` as ' + str(args_max_seq_length) + ' but '
'the maximum the model supports is ' + str(model_max_seq_length) + '. We shall reduce it.')
args.max_seq_length = model_max_seq_length
if model is not None and hasattr(model, 'for_training'):
model.for_training(use_gradient_checkpointing=getattr(args, 'gradient_checkpointing', True))
if 'tokenizer' in locals() and hasattr(tokenizer, 'padding_side'): tokenizer.padding_side = 'right'
if 'processing_class' in locals():
if hasattr(processing_class, 'padding_side'): processing_class.padding_side = 'right'
if hasattr(processing_class, 'tokenizer') and hasattr(processing_class.tokenizer, 'padding_side'): processing_class.tokenizer.padding_side = 'right'
other_metrics = []
if not isinstance(reward_funcs, list): _reward_funcs = [reward_funcs]
else: _reward_funcs = reward_funcs
for reward_func in _reward_funcs:
try:
reward_func_name = reward_func.__name__
if True:
other_metrics.append(f'rewards/{reward_func_name}/mean')
if True:
other_metrics.append(f'rewards/{reward_func_name}/std')
if False:
other_metrics.append(f'rewards/{reward_func_name}')
except: pass
from unsloth_zoo.logging_utils import PatchRLStatistics
PatchRLStatistics('grpo_trainer', other_metrics)
# [TODO] Fix up DataParallel multiplying batch sizes
# [TODO] DDP works, but DP seems to not work? [TODO]
if getattr(args, "parallel_mode", None) == ParallelMode.NOT_DISTRIBUTED and args.n_gpu > 1:
if getattr(args, "_n_gpu", 1) != 1:
args._n_gpu = 1
if "model" in locals() and hasattr(model, "for_training"):
model.for_training(use_gradient_checkpointing=getattr(args, 'gradient_checkpointing', True))
super().__init__(
model = model,
reward_funcs = reward_funcs,
args = args,
train_dataset = train_dataset,
eval_dataset = eval_dataset,
processing_class = processing_class,
reward_processing_classes = reward_processing_classes,
callbacks = callbacks,
peft_config = peft_config,**kwargs)
if "model" in locals() and hasattr(model, "for_inference"):
model.for_inference()
if hasattr(self, 'neftune_hook_handle'):
self.neftune_hook_handle.remove()
if hasattr(self, 'neftune_hook_handle'): del self.neftune_hook_handle
if getattr(args, 'neftune_noise_alpha', None) is not None:
model.get_input_embeddings().neftune_noise_alpha = self.neftune_noise_alpha
pass
if hasattr(self, 'accelerator'):
scaler = self.accelerator.scaler
current_model = model
while hasattr(current_model, 'model'):
current_model.accelerator_scaler = scaler
current_model = current_model.model
current_model.accelerator_scaler = scaler
pass
if hasattr(self, 'train'):
self.train = MethodType(prepare_for_training_mode(self.__class__.train), self)
pass
if hasattr(self, 'llm') and self.llm is not None and hasattr(self.llm, 'get_tokenizer'):
_vllm_tok = self.llm.get_tokenizer()
_pc = getattr(self, 'processing_class', None) or getattr(self, 'tokenizer', None)
if _vllm_tok is not None and _pc is not None and getattr(_pc, 'chat_template', None) is not None and getattr(_vllm_tok, 'chat_template', None) is None:
_vllm_tok.chat_template = _pc.chat_template
pass
pass
if hasattr(logger, "addFilter"):
import logging
class HideLoggingMessage(logging.Filter):
def __init__(self, text): self.text = text
def filter(self, x): return not (self.text in x.getMessage())
pass
logger.addFilter(HideLoggingMessage("`use_cache=True`"))