[None][fix] fix visual encoder attention weight loading for Qwen2.5-VL tp deployment (sm < 100) #8380
Conversation
Signed-off-by: Nekofish-L <[email protected]>
📝 WalkthroughWalkthroughAdds tensor-parallel sharding for Qwen2VL vision encoder QKV and output projection weights when running on SM < 100. Computes TP rank/size, head counts, shards qkv weights/biases and o_proj weights per rank, then loads the sharded state dict. Default loading remains unchanged otherwise. Changes
Sequence Diagram(s)sequenceDiagram
participant Loader as VisionWeightLoader
participant Env as RuntimeEnv
participant TP as TensorParallel
participant SD as StateDict
Loader->>Env: query SM version
Env-->>Loader: sm_version
Loader->>TP: get tp_rank, tp_size
alt SM < 100
Loader->>SD: fetch attn.qkv_proj (W,b)
Loader->>Loader: shard_qkv(W,b,num_heads,tp_size)
Loader->>SD: assign per-rank qkv shard (tp_rank)
Loader->>SD: fetch attn.o_proj.weight
Loader->>Loader: chunk by output dim
Loader->>SD: assign o_proj shard (tp_rank)
else SM >= 100
Loader->>SD: keep full weights (no sharding)
end
Loader->>Loader: load_state_dict(SD)
Loader-->>TP: continue with LLM loading
Estimated code review effort🎯 3 (Moderate) | ⏱️ ~25 minutes Pre-merge checks and finishing touches❌ Failed checks (2 warnings)
✅ Passed checks (1 passed)
✨ Finishing touches
🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 1
📜 Review details
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (1)
tensorrt_llm/_torch/models/modeling_qwen2vl.py(1 hunks)
🧰 Additional context used
📓 Path-based instructions (3)
**/*.{h,hpp,hh,hxx,cpp,cxx,cc,cu,cuh,py}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
Use only spaces, no tabs; indent with 4 spaces.
Files:
tensorrt_llm/_torch/models/modeling_qwen2vl.py
**/*.py
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
**/*.py: Python code must target Python 3.8+.
Indent Python code with 4 spaces; do not use tabs.
Maintain module namespace when importing; prefer 'from package.subpackage import foo' then 'foo.SomeClass()' instead of importing the class directly.
Python filenames should be snake_case (e.g., some_file.py).
Python classes use PascalCase names.
Functions and methods use snake_case names.
Local variables use snake_case; prefix 'k' for variables that start with a number (e.g., k_99th_percentile).
Global variables use upper SNAKE_CASE prefixed with 'G' (e.g., G_MY_GLOBAL).
Constants use upper SNAKE_CASE (e.g., MY_CONSTANT).
Avoid shadowing variables from an outer scope.
Initialize all externally visible members of a class in the constructor.
Prefer docstrings for interfaces that may be used outside a file; comments for in-function or file-local interfaces.
Use Google-style docstrings for classes and functions (Sphinx-parsable).
Document attributes and variables inline so they render under the class/function docstring.
Avoid reflection when a simpler, explicit approach suffices (e.g., avoid dict(**locals()) patterns).
In try/except, catch the most specific exceptions possible.
For duck-typing try/except, keep the try body minimal and use else for the main logic.
Files:
tensorrt_llm/_torch/models/modeling_qwen2vl.py
**/*.{cpp,cxx,cc,h,hpp,hh,hxx,cu,cuh,py}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
Prepend the NVIDIA Apache-2.0 copyright header with current year to the top of all source files (e.g., .cpp, .h, .cu, .py).
Files:
tensorrt_llm/_torch/models/modeling_qwen2vl.py
🧠 Learnings (2)
📚 Learning: 2025-09-29T15:14:28.503Z
Learnt from: amitz-nv
PR: NVIDIA/TensorRT-LLM#8063
File: tensorrt_llm/lora_manager.py:1080-1112
Timestamp: 2025-09-29T15:14:28.503Z
Learning: In tensorrt_llm/lora_manager.py, when calculating part_sizes for attn_qkv fused LoRA modules, the sizes are correctly multiplied by tp_size because model_config.num_heads and model_config.num_kv_heads are already divided by tp_size (per-TP-rank values), so multiplication is needed to get the original full concatenated dimension size. The interleave_fused_lora_weights_for_tp function provides proper validation.
Applied to files:
tensorrt_llm/_torch/models/modeling_qwen2vl.py
📚 Learning: 2025-09-29T15:14:28.503Z
Learnt from: amitz-nv
PR: NVIDIA/TensorRT-LLM#8063
File: tensorrt_llm/lora_manager.py:1080-1112
Timestamp: 2025-09-29T15:14:28.503Z
Learning: In tensorrt_llm/lora_manager.py, when calculating part_sizes for attn_qkv fused LoRA modules, the sizes are correctly multiplied by tp_size because model_config.num_heads and model_config.num_kv_heads are already divided by tp_size (per-TP-rank values), so multiplication is needed to get the original full concatenated dimension size. The interleave_fused_lora_weights_for_tp function provides proper validation with asserts for total size and TP divisibility.
Applied to files:
tensorrt_llm/_torch/models/modeling_qwen2vl.py
🧬 Code graph analysis (1)
tensorrt_llm/_torch/models/modeling_qwen2vl.py (2)
tensorrt_llm/_torch/distributed/communicator.py (1)
tp_size(63-64)tensorrt_llm/runtime/model_runner_cpp.py (1)
num_heads(505-506)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (1)
- GitHub Check: Pre-commit Check
| tp_rank = self.model_config.mapping.tp_rank | ||
| tp_size = self.model_config.mapping.tp_size | ||
| num_vision_heads = self.mm_encoder.config.num_heads | ||
|
|
||
| # Need to shard the weights to support tp | ||
| def shard_qkv(tensor, is_weight=True): | ||
| hidden_dim = tensor.shape[0] // 3 | ||
| head_dim = hidden_dim // num_vision_heads | ||
| heads_per_tp = num_vision_heads // tp_size | ||
| start, end = tp_rank * heads_per_tp, (tp_rank + | ||
| 1) * heads_per_tp | ||
|
|
||
| if is_weight: | ||
| tensor = tensor.reshape(3, num_vision_heads, head_dim, | ||
| hidden_dim) | ||
| sliced = tensor[:, start:end].reshape(3, -1, hidden_dim) | ||
| else: | ||
| tensor = tensor.reshape(3, num_vision_heads, head_dim) | ||
| sliced = tensor[:, start:end].reshape(3, -1) | ||
|
|
||
| return torch.cat([sliced[0], sliced[1], sliced[2]], dim=0) | ||
|
|
||
| for key in vision_weights.keys(): | ||
| if "attn.qkv_proj" in key: | ||
| if "weight" in key: | ||
| # qkv_proj.weight shape: [3 * hidden_dim, hidden_dim] | ||
| vision_weights[key] = shard_qkv(vision_weights[key], | ||
| is_weight=True) | ||
| elif "bias" in key: | ||
| # qkv_proj.bias shape: [3 * hidden_dim] | ||
| vision_weights[key] = shard_qkv(vision_weights[key], | ||
| is_weight=False) | ||
|
|
||
| if "attn.o_proj.weight" in key: | ||
| # o_proj.weight shape: [hidden_dim, hidden_dim] | ||
| vision_weights[key] = torch.chunk(vision_weights[key], | ||
| tp_size, |
There was a problem hiding this comment.
Add explicit head-count divisibility checks
heads_per_tp = num_vision_heads // tp_size silently floors. If someone configures TP with a head-count that isn’t cleanly divisible, we’ll reshape/slice to the wrong size and load_state_dict will fail (or worse, we’d drop heads). Please reject misconfigured checkpoints up front and validate that each per-projection slice splits cleanly across heads before reshaping.
num_vision_heads = self.mm_encoder.config.num_heads
# Need to shard the weights to support tp
def shard_qkv(tensor, is_weight=True):
- hidden_dim = tensor.shape[0] // 3
+ if tensor.shape[0] % 3 != 0:
+ raise ValueError(
+ f"Unexpected fused-qkv size {tensor.shape[0]} (not divisible by 3).")
+ hidden_dim = tensor.shape[0] // 3
+ if hidden_dim % num_vision_heads != 0:
+ raise ValueError(
+ f"Vision head split requires hidden_dim {hidden_dim} to be divisible by num_heads {num_vision_heads}.")
head_dim = hidden_dim // num_vision_heads
heads_per_tp = num_vision_heads // tp_size
start, end = tp_rank * heads_per_tp, (tp_rank +
1) * heads_per_tp
@@
- for key in vision_weights.keys():
+ if num_vision_heads % tp_size != 0:
+ raise ValueError(
+ f"Vision TP requires num_heads ({num_vision_heads}) to be divisible by tp_size ({tp_size}).")
+ for key in vision_weights.keys():📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
| tp_rank = self.model_config.mapping.tp_rank | |
| tp_size = self.model_config.mapping.tp_size | |
| num_vision_heads = self.mm_encoder.config.num_heads | |
| # Need to shard the weights to support tp | |
| def shard_qkv(tensor, is_weight=True): | |
| hidden_dim = tensor.shape[0] // 3 | |
| head_dim = hidden_dim // num_vision_heads | |
| heads_per_tp = num_vision_heads // tp_size | |
| start, end = tp_rank * heads_per_tp, (tp_rank + | |
| 1) * heads_per_tp | |
| if is_weight: | |
| tensor = tensor.reshape(3, num_vision_heads, head_dim, | |
| hidden_dim) | |
| sliced = tensor[:, start:end].reshape(3, -1, hidden_dim) | |
| else: | |
| tensor = tensor.reshape(3, num_vision_heads, head_dim) | |
| sliced = tensor[:, start:end].reshape(3, -1) | |
| return torch.cat([sliced[0], sliced[1], sliced[2]], dim=0) | |
| for key in vision_weights.keys(): | |
| if "attn.qkv_proj" in key: | |
| if "weight" in key: | |
| # qkv_proj.weight shape: [3 * hidden_dim, hidden_dim] | |
| vision_weights[key] = shard_qkv(vision_weights[key], | |
| is_weight=True) | |
| elif "bias" in key: | |
| # qkv_proj.bias shape: [3 * hidden_dim] | |
| vision_weights[key] = shard_qkv(vision_weights[key], | |
| is_weight=False) | |
| if "attn.o_proj.weight" in key: | |
| # o_proj.weight shape: [hidden_dim, hidden_dim] | |
| vision_weights[key] = torch.chunk(vision_weights[key], | |
| tp_size, | |
| tp_rank = self.model_config.mapping.tp_rank | |
| tp_size = self.model_config.mapping.tp_size | |
| num_vision_heads = self.mm_encoder.config.num_heads | |
| # Need to shard the weights to support tp | |
| def shard_qkv(tensor, is_weight=True): | |
| if tensor.shape[0] % 3 != 0: | |
| raise ValueError( | |
| f"Unexpected fused-qkv size {tensor.shape[0]} (not divisible by 3).") | |
| hidden_dim = tensor.shape[0] // 3 | |
| if hidden_dim % num_vision_heads != 0: | |
| raise ValueError( | |
| f"Vision head split requires hidden_dim {hidden_dim} to be divisible by num_heads {num_vision_heads}.") | |
| head_dim = hidden_dim // num_vision_heads | |
| heads_per_tp = num_vision_heads // tp_size | |
| start, end = tp_rank * heads_per_tp, (tp_rank + 1) * heads_per_tp | |
| if is_weight: | |
| tensor = tensor.reshape(3, num_vision_heads, head_dim, | |
| hidden_dim) | |
| sliced = tensor[:, start:end].reshape(3, -1, hidden_dim) | |
| else: | |
| tensor = tensor.reshape(3, num_vision_heads, head_dim) | |
| sliced = tensor[:, start:end].reshape(3, -1) | |
| return torch.cat([sliced[0], sliced[1], sliced[2]], dim=0) | |
| if num_vision_heads % tp_size != 0: | |
| raise ValueError( | |
| f"Vision TP requires num_heads ({num_vision_heads}) to be divisible by tp_size ({tp_size}).") | |
| for key in vision_weights.keys(): | |
| if "attn.qkv_proj" in key: | |
| if "weight" in key: | |
| # qkv_proj.weight shape: [3 * hidden_dim, hidden_dim] | |
| vision_weights[key] = shard_qkv(vision_weights[key], | |
| is_weight=True) | |
| elif "bias" in key: | |
| # qkv_proj.bias shape: [3 * hidden_dim] | |
| vision_weights[key] = shard_qkv(vision_weights[key], | |
| is_weight=False) | |
| if "attn.o_proj.weight" in key: | |
| # o_proj.weight shape: [hidden_dim, hidden_dim] | |
| vision_weights[key] = torch.chunk(vision_weights[key], | |
| tp_size, |
🤖 Prompt for AI Agents
In tensorrt_llm/_torch/models/modeling_qwen2vl.py around lines 1058 to 1094, add
explicit divisibility checks to reject misconfigured TP/head counts before
reshaping: validate that num_vision_heads % tp_size == 0 and raise a clear
ValueError if not; inside shard_qkv for weights assert tensor.shape[0] % 3 == 0,
compute hidden_dim = tensor.shape[0] // 3 then assert hidden_dim %
num_vision_heads == 0 before computing head_dim; for biases assert
tensor.shape[0] % 3 == 0 and (tensor.shape[0] // 3) % num_vision_heads == 0;
lastly, before torch.chunk(vision_weights[key], tp_size) check that
vision_weights[key].shape[0] % tp_size == 0 and raise a ValueError if not so
reshapes/chunks never silently floor or drop heads.
svc-trtllm-gh-bot
added
the
Community want to contribute
|
Thx for your PR, @Nekofish-L. I think I missed TP > 1 part when doing refactoring. |
|
/bot run |
|
PR_Github #22018 [ run ] triggered by Bot. Commit: |
|
PR_Github #22018 [ run ] completed with state |
|
@yechank-nvidia Thanks for your review! |
|
@Nekofish-L I see. Then, I will not merge this PR but rather trying to come-up with the proper fix. In the meantime, I will leave this PR as a reference for others who is facing weight-loading issue for Qwen2.5-VL weight loading when TP > 1. |
|
Hi @Nekofish-L, I submitted PR for the weight loading fix: #8680. You can try this PR or wait to be merged into the main. Once it is merged to the main, I will close this PR. Thx. |
yechank-nvidia
added
the
Multimodal
|
Closing this for now as #8680 is merged to the main. Thx for the contribution. |
4 participants
Summary by CodeRabbit
Description
This PR fixes an issue where, during tensor parallel deployment of Qwen2.5-VL(sm < 100), each rank incorrectly received the full, un-split visual encoder attention weights instead of their respective shards.
Test Coverage
PR Checklist
Please review the following before submitting your PR:
PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.
PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.
Test cases are provided for new code paths (see test instructions)
Any new dependencies have been scanned for license and vulnerabilities
CODEOWNERS updated if ownership changes
Documentation updated as needed
The reviewers assigned automatically/manually are appropriate for the PR.
Please check this after reviewing the above items as appropriate for this PR.
GitHub Bot Help
/bot [-h] ['run', 'kill', 'skip', 'reuse-pipeline'] ...Provide a user friendly way for developers to interact with a Jenkins server.
Run
/bot [-h|--help]to print this help message.See details below for each supported subcommand.
run [--reuse-test (optional)pipeline-id --disable-fail-fast --skip-test --stage-list "A10-PyTorch-1, xxx" --gpu-type "A30, H100_PCIe" --test-backend "pytorch, cpp" --add-multi-gpu-test --only-multi-gpu-test --disable-multi-gpu-test --post-merge --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx" --detailed-log --debug(experimental)]Launch build/test pipelines. All previously running jobs will be killed.
--reuse-test (optional)pipeline-id(OPTIONAL) : Allow the new pipeline to reuse build artifacts and skip successful test stages from a specified pipeline or the last pipeline if no pipeline-id is indicated. If the Git commit ID has changed, this option will be always ignored. The DEFAULT behavior of the bot is to reuse build artifacts and successful test results from the last pipeline.--disable-reuse-test(OPTIONAL) : Explicitly prevent the pipeline from reusing build artifacts and skipping successful test stages from a previous pipeline. Ensure that all builds and tests are run regardless of previous successes.--disable-fail-fast(OPTIONAL) : Disable fail fast on build/tests/infra failures.--skip-test(OPTIONAL) : Skip all test stages, but still run build stages, package stages and sanity check stages. Note: Does NOT update GitHub check status.--stage-list "A10-PyTorch-1, xxx"(OPTIONAL) : Only run the specified test stages. Examples: "A10-PyTorch-1, xxx". Note: Does NOT update GitHub check status.--gpu-type "A30, H100_PCIe"(OPTIONAL) : Only run the test stages on the specified GPU types. Examples: "A30, H100_PCIe". Note: Does NOT update GitHub check status.--test-backend "pytorch, cpp"(OPTIONAL) : Skip test stages which don't match the specified backends. Only support [pytorch, cpp, tensorrt, triton]. Examples: "pytorch, cpp" (does not run test stages with tensorrt or triton backend). Note: Does NOT update GitHub pipeline status.--only-multi-gpu-test(OPTIONAL) : Only run the multi-GPU tests. Note: Does NOT update GitHub check status.--disable-multi-gpu-test(OPTIONAL) : Disable the multi-GPU tests. Note: Does NOT update GitHub check status.--add-multi-gpu-test(OPTIONAL) : Force run the multi-GPU tests in addition to running L0 pre-merge pipeline.--post-merge(OPTIONAL) : Run the L0 post-merge pipeline instead of the ordinary L0 pre-merge pipeline.--extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx"(OPTIONAL) : Run the ordinary L0 pre-merge pipeline and specified test stages. Examples: --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx".--detailed-log(OPTIONAL) : Enable flushing out all logs to the Jenkins console. This will significantly increase the log volume and may slow down the job.--debug(OPTIONAL) : Experimental feature. Enable access to the CI container for debugging purpose. Note: Specify exactly one stage in thestage-listparameter to access the appropriate container environment. Note: Does NOT update GitHub check status.For guidance on mapping tests to stage names, see
docs/source/reference/ci-overview.mdand the
scripts/test_to_stage_mapping.pyhelper.kill
killKill all running builds associated with pull request.
skip
skip --comment COMMENTSkip testing for latest commit on pull request.
--comment "Reason for skipping build/test"is required. IMPORTANT NOTE: This is dangerous since lack of user care and validation can cause top of tree to break.reuse-pipeline
reuse-pipelineReuse a previous pipeline to validate current commit. This action will also kill all currently running builds associated with the pull request. IMPORTANT NOTE: This is dangerous since lack of user care and validation can cause top of tree to break.