Upgrade trl fix #2544
Merged
Upgrade trl fix #2544
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…se autoSequenceClassification
Trl upgrade fix
For TRL 0.18.0 (Main branch of TRL at the time because its on 0.17.0) , the SFT trainer for some reason deletes the labels column and unsloth internal loss funcitons need that column for hte claculations so I add it back in like this.
Trl update, particualry small fixes on SFT trainer
Signed-off-by: datta0 <[email protected]>
unsloth/models/rl.py
Outdated
| if trl_version >= "0.18": | ||
| # Replace LLM init with already existing vLLM engine for colocate mode | ||
| vllm_llm_init_pattern = r"self\.llm\s*=\s*LLM\([^)]*\)*\)" | ||
| vllm_llm_repalcement = "self.llm = model.vllm_engine\n" |
unsloth/models/rl.py
Outdated
| vllm_llm_repalcement = "self.llm = model.vllm_engine\n" | ||
| new_vllm_part = re.sub( | ||
| vllm_llm_init_pattern, | ||
| vllm_llm_repalcement, |
unsloth/models/rl_replacements.py
Outdated
| 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 | ||
| os.environ["UNSLOTH_RETURN_HIDDEN_STATES"] = "0" |
There was a problem hiding this comment.
Wait this forcefully returns hidden states right - doesn't this make the GRPO loss use more memory?
unsloth/models/rl_replacements.py
Outdated
| logits = model(input_ids=input_ids, attention_mask=attention_mask, logits_to_keep=logits_to_keep + 1).logits | ||
| logits = logits[:, :-1, :] # (B, L-1, V), exclude the last logit: it corresponds to the next token pred | ||
|
|
||
| logits = logits.to(torch.float32) |
There was a problem hiding this comment.
I don't think this is correct since we auto upcast inside the torch.compile function, so this uses 2x more memory
There was a problem hiding this comment.
yeah I think so.
will remove this
unsloth/models/rl_replacements.py
Outdated
| logits = logits[:, -logits_to_keep:] | ||
| return logits | ||
| # return selective_log_softmax(logits, input_ids) # compute logprobs for the input tokens | ||
| #return logits |
There was a problem hiding this comment.
Yes this is correct for non GRPO, but reminder in Unsloth GRPO, we specifically calculate the logits on the fly to save VRAM
Signed-off-by: datta0 <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Depends on : unslothai/unsloth-zoo#140
Tested: GRPO with qwen-3-4B, ORPO and DPO (to make sure nothing breaks there)
TODO:
[ ] Compare performance