[GPU] Fix access dimension error in dynamic shape #32626
Merged
+56
−42
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
The regression was observed in the recent memory reset bugfix PR. The issue is get_dims() failure that occurs when calling the convolution feature in dynamic shape. Solution: The fix integrates the logic from the existing get_conv_channel_count() utility to correctly and safely determine the channel dimension of the convolution's input/output under dynamic condition, preventing the dimension access failure. Signed-off-by: hyunback <[email protected]>
hyunback
added
category: GPU
hyunback
changed the title
|
Looks like this is a fix for the same problem #32601 |
isanghao
reviewed
Oct 31, 2025
| // If the channel count is dynamic, we cannot verify feature alignment, | ||
| // so we conservatively skip the reset and return false for this condition. | ||
| if (in_channel_count == -1) | ||
| return false; |
There was a problem hiding this comment.
I think the conservative behavior here is to reset memory for potential accuracy issue. Do we have a case where this makes difference in performance? If not, let's return true.
There was a problem hiding this comment.
Agree, return true is correct. Applied
isanghao
reviewed
Oct 31, 2025
| // if layout is single blocked and feature size is not aligned with the blocking size, need to reset output so that we can guarantee zero-filling | ||
| // NOTE: We may improve this logic to avoid reset if we are sure that it is not "corrupted" by other layers. | ||
| if (output_layout.feature() % feature_block_size != 0) { | ||
| if (in_channel_count % feature_block_size != 0) { |
There was a problem hiding this comment.
could you add test case in mem_reset_test.cpp?
Signed-off-by: hyunback <[email protected]>
github-merge-queue bot
pushed a commit
that referenced
this pull request
Nov 3, 2025
### Description of the issue(symptom, root-cause, how it was resolved) The regression was observed in the recent memory reset bugfix PR. The issue is get_dims() failure that occurs when calling the convolution feature in dynamic shape. Solution: The fix integrates the logic from the existing get_conv_channel_count() utility to correctly and safely determine the channel dimension of the convolution's input/output under dynamic condition, preventing the dimension access failure. #### Reproduction step and snapshot (if applicable. Do not attach for customer model) ##### E2E python tools/llm_bench/benchmark.py -m ov-share-05.sclab.intel.com/cv_bench_cache/latest_models_llm/qwen2-vl-7b-instruct/pytorch/ov/OV_FP16-4BIT_DEFAULT -d GPU.1 -mc 1 -ic 256 -n 3 -pf frameworks.ai.openvino.llm.prompts/32_1024/qwen2-vl-7b-instruct.jsonl ###### Benchmark_app benchmark_app -d GPU.1 --hint none -nireq 1 -niter 1 -m ov-share-05.sclab.intel.com/cv_bench_cache/latest_models_llm/qwen2-vl-7b-instruct/pytorch/ov/OV_FP16-4BIT_DEFAULT/openvino_vision_embeddings_model.xml -data_shape hidden_states[1,1176] #### Checklist - [x] Is it a proper fix? Yes - [x] Did you include test case for this fix, if necessary? - [x] Did you review existing test that can be extended to cover this scenario? Which test did you review? mem_reset_test.cpp ### Tickets: - *CVS-175613* --------- Signed-off-by: hyunback <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 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.
Description of the issue(symptom, root-cause, how it was resolved)
The regression was observed in the recent memory reset bugfix PR. The issue is get_dims() failure that occurs when calling the convolution feature in dynamic shape.
Solution:
The fix integrates the logic from the existing get_conv_channel_count() utility to correctly and safely determine the channel dimension of the convolution's input/output under dynamic condition, preventing the dimension access failure.
Reproduction step and snapshot (if applicable. Do not attach for customer model)
E2E
python tools/llm_bench/benchmark.py -m ov-share-05.sclab.intel.com/cv_bench_cache/latest_models_llm/qwen2-vl-7b-instruct/pytorch/ov/OV_FP16-4BIT_DEFAULT -d GPU.1 -mc 1 -ic 256 -n 3 -pf frameworks.ai.openvino.llm.prompts/32_1024/qwen2-vl-7b-instruct.jsonl
Benchmark_app
benchmark_app -d GPU.1 --hint none -nireq 1 -niter 1 -m ov-share-05.sclab.intel.com/cv_bench_cache/latest_models_llm/qwen2-vl-7b-instruct/pytorch/ov/OV_FP16-4BIT_DEFAULT/openvino_vision_embeddings_model.xml -data_shape hidden_states[1,1176]
Checklist
Tickets: