Fix gpt-j-6b RTN RuntimeError#1848
Merged
Merged
Conversation
Signed-off-by: lvliang-intel <liang1.lv@intel.com>
…ix_gptj_dtype_issue
for more information, see https://pre-commit.ci
Contributor
There was a problem hiding this comment.
Pull request overview
Fixes the GPT-J-6B RTN runtime error caused by missing / incorrectly-typed position_ids during cached-input replay by improving how block inputs are captured, normalized, and replayed.
Changes:
- Wraps calibration block-forward hooks to convert positional args into named kwargs for caching.
- Adds input preprocessing to unwrap single-element containers and preserve integer tensor dtypes.
- Updates
block_forwardto replay blocks using keyword arguments and normalizesposition_idswhen needed.
Reviewed changes
Copilot reviewed 4 out of 4 changed files in this pull request and generated 4 comments.
| File | Description |
|---|---|
| auto_round/compressors/utils.py | Reworks block replay to map positional inputs to real parameter names and call blocks via kwargs; adds position_ids normalization. |
| auto_round/compressors/data_driven.py | Adjusts cached-input dtype/device preparation, including unwrapping single-element containers and list handling. |
| auto_round/calibration/inputs.py | Adds shared preprocessing helper to unwrap single-element kwargs and preserves int index tensors from dtype casting. |
| auto_round/calibration/hooks.py | Wraps block-forward capture hook with positional→kwargs conversion and updates replay invocation. |
| kwargs.pop("hidden_states") | ||
| return m.orig_forward(hidden_states, *positional_inputs, **kwargs) | ||
| kwargs.pop("hidden_states", None) | ||
| return m.orig_forward(hidden_states=hidden_states, *positional_inputs, **kwargs) |
Comment on lines
+180
to
+197
| # Use the block's actual parameter name for the first positional argument. | ||
| import inspect as _inspect | ||
|
|
||
| param_names = [p for p in _inspect.signature(block.forward).parameters.keys() if p != "self"] | ||
| block_input_kwarg = param_names[0] if param_names else "hidden_states" | ||
| if block_input_kwarg not in input_others: | ||
| input_others[block_input_kwarg] = input_ids | ||
|
|
||
| # Convert positional inputs to keyword args for any remaining positional parameters. | ||
| positional_inputs = input_tuple or () | ||
| if positional_inputs: | ||
| for i, val in enumerate(positional_inputs): | ||
| param_idx = i + 1 # hidden_states is params[0] | ||
| if param_idx < len(param_names): | ||
| param_name = param_names[param_idx] | ||
| if param_name not in input_others: | ||
| input_others[param_name] = val | ||
| positional_inputs = () |
Comment on lines
+977
to
+979
| to_dtype(v, tmp_dtype) | ||
| for v in val | ||
| if not (isinstance(v, torch.Tensor) and v.dtype in (torch.int32, torch.int64)) |
Comment on lines
+199
to
+219
| # Guard: ensure position_ids is a tensor, not a list or None. | ||
| if "position_ids" in input_others: | ||
| pid = input_others["position_ids"] | ||
| if isinstance(pid, list): | ||
| if len(pid) == 1: | ||
| input_others["position_ids"] = pid[0] | ||
| elif len(pid) == 0: | ||
| # Generate position_ids from hidden_states shape when it's empty. | ||
| input_others["position_ids"] = ( | ||
| torch.arange(input_ids.shape[1], device=input_ids.device, dtype=torch.long) | ||
| .unsqueeze(0) | ||
| .expand(input_ids.shape[0], -1) | ||
| ) | ||
| elif pid is None: | ||
| # Generate position_ids from hidden_states shape when it's None. | ||
| input_others["position_ids"] = ( | ||
| torch.arange(input_ids.shape[1], device=input_ids.device, dtype=torch.long) | ||
| .unsqueeze(0) | ||
| .expand(input_ids.shape[0], -1) | ||
| ) | ||
|
|
…ix_gptj_dtype_issue
…uto-round into lvl/fix_gptj_dtype_issue
Contributor
Author
|
CUDA_VISIBLE_DEVICES=4 auto-round --model_name /mnt/disk3/lvl/gpt-j-6b/ --bits 4 --iters 0 --tasks lambada_openai Notes:
evaluation running time=95s |
Contributor
Author
|
/azp run Unit-Test-CUDA-AutoRound |
|
Azure Pipelines successfully started running 1 pipeline(s). |
Signed-off-by: lvliang-intel <liang1.lv@intel.com>
Contributor
Author
|
/azp run Unit-Test-CUDA-AutoRound |
|
Azure Pipelines successfully started running 1 pipeline(s). |
XuehaoSun
approved these changes
May 25, 2026
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Description
Fix gpt-j-6b RTN RuntimeError
root cause:
GPT-J calls block(hidden_states, position_ids, attention_mask), position_ids is a positional arg, not a keyword arg. The original hook only captured keyword args, so position_ids was never stored in state.inputs. On replay, position_ids became None or a Python list instead of a torch.Tensor, and torch.gather requires int32/int64, caused the error.
Solution:
The hook wasn't capturing positional args like GPT-J's position_ids, so it was missing on replay. The fix wraps the hook to intercept positional args as keywords, unwraps single-element containers from batch_size=1, and invokes blocks via block(**input_others) using real parameter names.
Type of Change
Bug fix
Related Issues
#1838
Checklist Before Submitting
/azp run Unit-Test-CUDA-AutoRound.