Enabled and optimized GLM-4v-9b on Gaudi #691
Merged
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
gyou2021
requested review from
afierka-intel,
kzawora-intel,
madamczyk-intel,
mgawarkiewicz,
michalkuligowski and
vivekgoe
as code owners
jikunshang
reviewed
Jan 26, 2025
PatrykWo
added
New Model
|
@gyou2021 Recently vllm has some update about multi-modal support, especially they introduced InputProcessor for each model, like [VLM] Implement merged multimodal processor for Mllama. Will it affect your implementation and will you do a update based on this? |
Author
No. It won't affect the implementation. So no update based on this is needed. |
michalkuligowski
requested changes
Mar 21, 2025
michalkuligowski
requested changes
Mar 24, 2025
michalkuligowski
requested changes
Mar 27, 2025
michalkuligowski
approved these changes
Mar 28, 2025
Signed-off-by: gyou2021 <ganmei.you@intel.com>
Signed-off-by: gyou2021 <ganmei.you@intel.com>
Signed-off-by: gyou2021 <ganmei.you@intel.com>
Signed-off-by: gyou2021 <ganmei.you@intel.com>
|
/run-gaudi-tests |
|
/run-gaudi-tests |
imangohari1
added a commit
to imangohari1/vllm-fork
that referenced
this pull request
Apr 8, 2025
This reverts commit c0e696b.
michalkuligowski
added a commit
that referenced
this pull request
Apr 16, 2025
This PR reverts #691 that leads to `AttributeError: 'tuple' object has no attribute 'reshape'` for Qwen2.5-VL. ## Test server: ``` python -m vllm.entrypoints.openai.api_server --port 8080 --model Qwen/Qwen2.5-VL-3B-Instruct --tensor-parallel-size 1 --max-num-seqs 128 --dtype bfloat16 --gpu-memory-util 0.9 --max-num-batched-tokens 32768 --max-model-len 32768 --block-size 128 ``` Client: ``` python benchmark_serving.py --backend openai-chat --model Qwen/Qwen2.5-VL-3B-Instruct --trust-remote-code --port 8080 --endpoint /v1/chat/completions --dataset-path lmarena-ai/vision-arena-bench-v0.1 --dataset-name hf --hf-split train --num-prompts 40 --request-rate inf --seed 0 --ignore_eos ``` ### as is ``` ERROR 04-14 16:51:14 engine.py:139] AttributeError("'tuple' object has no attribute 'reshape'") ERROR 04-14 16:51:14 engine.py:139] Traceback (most recent call last): ERROR 04-14 16:51:14 engine.py:139] File "/root/vllm-fork/vllm/engine/multiprocessing/engine.py", line 137, in start ERROR 04-14 16:51:14 engine.py:139] self.run_engine_loop() ERROR 04-14 16:51:14 engine.py:139] File "/root/vllm-fork/vllm/engine/multiprocessing/engine.py", line 200, in run_engine_loop ERROR 04-14 16:51:14 engine.py:139] request_outputs = self.engine_step() ERROR 04-14 16:51:14 engine.py:139] File "/root/vllm-fork/vllm/engine/multiprocessing/engine.py", line 218, in engine_step ERROR 04-14 16:51:14 engine.py:139] raise e ERROR 04-14 16:51:14 engine.py:139] File "/root/vllm-fork/vllm/engine/multiprocessing/engine.py", line 209, in engine_step ERROR 04-14 16:51:14 engine.py:139] return self.engine.step() ERROR 04-14 16:51:14 engine.py:139] File "/root/vllm-fork/vllm/engine/llm_engine.py", line 1380, in step ERROR 04-14 16:51:14 engine.py:139] outputs = self.model_executor.execute_model( ERROR 04-14 16:51:14 engine.py:139] File "/root/vllm-fork/vllm/executor/executor_base.py", line 138, in execute_model ERROR 04-14 16:51:14 engine.py:139] output = self.collective_rpc("execute_model", ERROR 04-14 16:51:14 engine.py:139] File "/root/vllm-fork/vllm/executor/uniproc_executor.py", line 58, in collective_rpc ERROR 04-14 16:51:14 engine.py:139] answer = run_method(self.driver_worker, method, args, kwargs) ERROR 04-14 16:51:14 engine.py:139] File "/root/vllm-fork/vllm/utils.py", line 2323, in run_method ERROR 04-14 16:51:14 engine.py:139] return func(*args, **kwargs) ERROR 04-14 16:51:14 engine.py:139] File "/root/vllm-fork/vllm/worker/hpu_worker.py", line 294, in execute_model ERROR 04-14 16:51:14 engine.py:139] output = LocalOrDistributedWorkerBase.execute_model( ERROR 04-14 16:51:14 engine.py:139] File "/root/vllm-fork/vllm/worker/worker_base.py", line 418, in execute_model ERROR 04-14 16:51:14 engine.py:139] output = self.model_runner.execute_model( ERROR 04-14 16:51:14 engine.py:139] File "/usr/local/lib/python3.10/dist-packages/torch/utils/_contextlib.py", line 116, in decorate_context ERROR 04-14 16:51:14 engine.py:139] return func(*args, **kwargs) ERROR 04-14 16:51:14 engine.py:139] File "/root/vllm-fork/vllm/worker/hpu_model_runner.py", line 2697, in execute_model ERROR 04-14 16:51:14 engine.py:139] hidden_states = self.model.forward( ERROR 04-14 16:51:14 engine.py:139] File "/usr/local/lib/python3.10/dist-packages/habana_frameworks/torch/hpu/graphs.py", line 745, in forward ERROR 04-14 16:51:14 engine.py:139] return wrapped_hpugraph_forward( ERROR 04-14 16:51:14 engine.py:139] File "/usr/local/lib/python3.10/dist-packages/habana_frameworks/torch/hpu/graphs.py", line 610, in wrapped_hpugraph_forward ERROR 04-14 16:51:14 engine.py:139] outputs = orig_fwd(*args, **kwargs) ERROR 04-14 16:51:14 engine.py:139] File "/root/vllm-fork/vllm/worker/hpu_model_runner.py", line 423, in forward ERROR 04-14 16:51:14 engine.py:139] hidden_states = self.model(*args, **kwargs) ERROR 04-14 16:51:14 engine.py:139] File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1742, in _wrapped_call_impl ERROR 04-14 16:51:14 engine.py:139] return self._call_impl(*args, **kwargs) ERROR 04-14 16:51:14 engine.py:139] File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1848, in _call_impl ERROR 04-14 16:51:14 engine.py:139] return inner() ERROR 04-14 16:51:14 engine.py:139] File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1796, in inner ERROR 04-14 16:51:14 engine.py:139] result = forward_call(*args, **kwargs) ERROR 04-14 16:51:14 engine.py:139] File "/root/vllm-fork/vllm/model_executor/models/qwen2_5_vl.py", line 1104, in forward ERROR 04-14 16:51:14 engine.py:139] inputs_embeds = self.get_input_embeddings_v0( ERROR 04-14 16:51:14 engine.py:139] File "/root/vllm-fork/vllm/model_executor/models/qwen2_5_vl.py", line 1037, in get_input_embeddings_v0 ERROR 04-14 16:51:14 engine.py:139] inputs_embeds = merge_multimodal_embeddings( ERROR 04-14 16:51:14 engine.py:139] File "/root/vllm-fork/vllm/model_executor/models/utils.py", line 448, in merge_multimodal_embeddings ERROR 04-14 16:51:14 engine.py:139] return _hpu_merge_multimodal_embeddings( ERROR 04-14 16:51:14 engine.py:139] File "/root/vllm-fork/vllm/model_executor/models/utils.py", line 674, in _hpu_merge_multimodal_embeddings ERROR 04-14 16:51:14 engine.py:139] multimodal_embeddings = multimodal_embeddings.reshape(-1, hidden_size) ``` ### with this PR ``` 100%|██████████| 1/1 [00:01<00:00, 1.14s/it] ============ Serving Benchmark Result ============ Successful requests: 1 Benchmark duration (s): 1.14 Total input tokens: 52 Total generated tokens: 128 Request throughput (req/s): 0.88 Output token throughput (tok/s): 112.62 Total Token throughput (tok/s): 158.37 ---------------Time to First Token---------------- Mean TTFT (ms): 169.75 Median TTFT (ms): 169.75 P99 TTFT (ms): 169.75 -----Time per Output Token (excl. 1st token)------ Mean TPOT (ms): 7.61 Median TPOT (ms): 7.61 P99 TPOT (ms): 7.61 ---------------Inter-token Latency---------------- Mean ITL (ms): 7.55 Median ITL (ms): 7.59 P99 ITL (ms): 8.10 ================================================== ``` Co-authored-by: Michał Kuligowski <mkuligowski@habana.ai>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 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.
Removed graph recompile caused by the image and different batch size;
Example:
python examples/offline_inference/vision_language.py -m glm4v