Selective Log softmax

unsloth/models/rl.py

@@ -24,11 +24,13 @@ import re
 import torch
 from unsloth_zoo.compiler import create_new_function
 from unsloth_zoo.logging_utils import PatchRLStatistics
+from unsloth_zoo.rl_replacements import RL_REPLACEMENTS
 from .rl_replacements import (
     RL_EXTRA_ARGS,
     RL_FUNCTIONS,
     RL_PRE_ITEMS,
 )
+selective_log_softmax = RL_REPLACEMENTS["selective_log_softmax"]
 
 torch_compile_options = {
     "epilogue_fusion"   : True,
@@ -84,19 +86,6 @@ def PatchRL(FastLanguageModel):
 pass
 
 
-# https://github.com/huggingface/trl/blob/main/trl/trainer/utils.py#L1674
-@torch.compile(dynamic = True, fullgraph = True, options = torch_compile_options,)
-def selective_log_softmax(logits, index):
-    logits = logits.to(torch.float32)
-    selected_logits = torch.gather(logits, dim=-1, index=index.unsqueeze(-1)).squeeze(-1)
-    # loop to reduce peak mem consumption
-    # logsumexp_values = torch.stack([torch.logsumexp(lg, dim=-1) for lg in logits])
-    logsumexp_values = torch.logsumexp(logits, dim = -1)
-    per_token_logps = selected_logits - logsumexp_values  # log_softmax(x_i) = x_i - logsumexp(x)
-    return per_token_logps
-pass
-
-
 RLTrainer_replacement = '''
 import os
 from typing import *
@@ -420,7 +409,7 @@ def _patch_trl_rl_trainers(trainer_file = "grpo_trainer"):
 
     # Selective log softmax
     selective_log_softmax_code = inspect.getsource(selective_log_softmax)
-    
+
     # Get final source code
     RLTrainer_source = RLTrainer_replacement.format(
         RLTrainer_name       = RLTrainer_name,

unsloth/models/rl_replacements.py

@@ -22,6 +22,7 @@ import re
 import torch
 import inspect
 from collections import defaultdict
+from unsloth_zoo.rl_replacements import RL_REPLACEMENTS
 RL_EXTRA_ARGS = defaultdict(list)
 RL_FUNCTIONS  = defaultdict(list)
 RL_PRE_ITEMS  = defaultdict(list)
@@ -193,45 +194,14 @@ def grpo_trainer__get_per_token_logps(function_name, function):
 pass
 RL_FUNCTIONS["grpo_trainer"].append(grpo_trainer__get_per_token_logps)
 
-
-# Custom compiled GRPO loss - creates 3 Triton kernels
-@torch.compile(dynamic = True, fullgraph = True, options = torch_compile_options,)
-def grpo_compute_loss(old_logits, new_logits, input_ids, mask, beta, advantages):
-    old_logits = old_logits.to(torch.float32)
-    new_logits = new_logits.to(torch.float32)
-    input_ids  = input_ids.unsqueeze(-1)
-
-    # x_i - logsumexp(x_i)
-    old_x = torch.gather(old_logits, dim = -1, index = input_ids).squeeze(-1)
-    new_x = torch.gather(new_logits, dim = -1, index = input_ids).squeeze(-1)
-    old = old_x - torch.logsumexp(old_logits, dim = -1)
-    new = new_x - torch.logsumexp(new_logits, dim = -1)
-
-    kl_i = torch.exp(old - new) - (old - new) - 1.0
-    loss_i = torch.exp(new - new.detach()) * advantages.unsqueeze(1)
-    loss_i = -(loss_i - beta * kl_i)
-
-    mask = mask.to(torch.float32)
-    n_mask_per_reward = mask.sum(1)
-    loss_per_reward = (loss_i * mask).sum(1) / n_mask_per_reward
-    loss = loss_per_reward.mean()
-    
-    # Get metrics as well which are folded
-    with torch.inference_mode():
-        completion_length = n_mask_per_reward.mean()
-        mean_kl_per_reward = (kl_i * mask).sum(1) / n_mask_per_reward
-        mean_kl = mean_kl_per_reward.mean()
-    pass
-    return loss, completion_length, mean_kl
-pass
-RL_PRE_ITEMS["grpo_trainer"].append(inspect.getsource((grpo_compute_loss)))
-
+grpo_compute_loss = RL_REPLACEMENTS["grpo_compute_loss"]
+RL_PRE_ITEMS["grpo_trainer"].append(inspect.getsource(grpo_compute_loss))
 
 # Edit _get_per_token_logps to handle mixed precision
 def grpo_trainer_compute_loss(function_name, function):
     if  function_name != "compute_loss": return function
 
-    def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=None):
+    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