[None][autodeploy] small refactors on attention matching #8079
Conversation
…ention_layout to use pattern matcher Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
Fridah-nv
force-pushed
the
user/fridah/attn-layout
branch
from
16d3071 to
eee5637
Compare
Fridah-nv
changed the title
Signed-off-by: Fridah-nv <201670829+Fridah-nv@users.noreply.github.com>
Signed-off-by: Fridah-nv <201670829+Fridah-nv@users.noreply.github.com>
Fridah-nv
requested review from
MrGeva and
lucaslie
and removed request for
MrGeva
📝 WalkthroughWalkthroughUnified attention custom op to auto_deploy::torch_attention with explicit layout handling ("bnsd" or "bsnd"), removed legacy grouped SDPA variants, updated backends and model patches to reference the new op, introduced dynamic generation/registration of attention transform patterns, adjusted sharding/KV-cache transformers, and updated tests and README accordingly. Changes
Sequence Diagram(s)sequenceDiagram
autonumber
actor Caller as Model/Module
participant Op as torch.ops.auto_deploy.torch_attention
participant Impl as Attention Impl
participant Kernel as Grouped SDPA Kernel
Caller->>Op: torch_attention(Q,K,V, ..., layout)
Note over Op: Validate layout in {"bnsd","bsnd"}
alt layout == "bsnd"
Op->>Impl: Transpose inputs to internal bnsd
else layout == "bnsd"
Op->>Impl: Use inputs as-is
end
Impl->>Kernel: Compute attention (grouped SDPA)
alt layout == "bsnd"
Impl->>Op: Transpose output back to bsnd
else
Impl-->>Op: Return bnsd output
end
Op-->>Caller: Output tensor
sequenceDiagram
autonumber
participant Gen as Pattern Generators
participant Reg as ADPattern Registry
participant GM as GraphModule
participant Pass as Transform Pass
Gen->>Reg: Generate grouped-attn patterns
Gen->>Reg: Generate layout patterns (bnsd/bsnd)
Pass->>GM: Apply registered patterns
alt Matches found
GM-->>Pass: Replacements applied
else No matches
GM-->>Pass: No-op
end
Estimated code review effort🎯 4 (Complex) | ⏱️ ~60 minutes Pre-merge checks and finishing touches❌ Failed checks (2 warnings, 1 inconclusive)
✨ Finishing touches
🧪 Generate unit tests
📜 Recent review detailsConfiguration used: Path: .coderabbit.yaml Review profile: CHILL Plan: Pro 📒 Files selected for processing (13)
🧰 Additional context used📓 Path-based instructions (3)**/*.{h,hpp,hh,hxx,cpp,cxx,cc,cu,cuh,py}📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
Files:
**/*.py📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
Files:
**/*.{cpp,cxx,cc,h,hpp,hh,hxx,cu,cuh,py}📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
Files:
🧬 Code graph analysis (10)tensorrt_llm/_torch/auto_deploy/custom_ops/torch_backend_attention.py (1)
tensorrt_llm/_torch/auto_deploy/models/patches/gptoss.py (1)
tensorrt_llm/_torch/auto_deploy/custom_ops/triton_attention.py (1)
tensorrt_llm/_torch/auto_deploy/transform/library/sharding.py (1)
tests/unittest/_torch/auto_deploy/unit/singlegpu/transformations/library/test_kv_cache.py (1)
tensorrt_llm/_torch/auto_deploy/transform/library/kvcache_transformers.py (1)
tests/unittest/_torch/auto_deploy/unit/multigpu/transformations/library/test_tp_sharding.py (1)
tests/unittest/_torch/auto_deploy/unit/singlegpu/transformations/library/test_attention_matcher_hf.py (2)
tests/unittest/_torch/auto_deploy/unit/singlegpu/transformations/library/test_attention_matcher.py (3)
tensorrt_llm/_torch/auto_deploy/transform/library/attention.py (2)
🪛 Ruff (0.13.1)tensorrt_llm/_torch/auto_deploy/custom_ops/torch_attention.py119-119: Avoid specifying long messages outside the exception class (TRY003) 218-218: Unused function argument: (ARG001) 220-220: Unused function argument: (ARG001) 221-221: Unused function argument: (ARG001) 222-222: Unused function argument: (ARG001) 223-223: Unused function argument: (ARG001) 224-224: Unused function argument: (ARG001) 225-225: Unused function argument: (ARG001) 226-226: Unused function argument: (ARG001) 227-227: Unused function argument: (ARG001) tensorrt_llm/_torch/auto_deploy/transform/library/attention.py389-389: Use explicit conversion flag Replace with conversion flag (RUF010) 402-402: Use explicit conversion flag Replace with conversion flag (RUF010) 426-426: Use of (S102) 443-443: Use of (S102) 512-512: Avoid specifying long messages outside the exception class (TRY003) 600-600: Use explicit conversion flag Replace with conversion flag (RUF010) 622-622: Use of (S102) 643-643: Use of (S102) 699-699: Avoid specifying long messages outside the exception class (TRY003) 754-754: Avoid specifying long messages outside the exception class (TRY003) ⏰ 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)
🔇 Additional comments (23)
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 |
There was a problem hiding this comment.
I migrated MatchGroupedAttention and MatchAttentionLayout to auto generate and register patterns, but I leave MatchEagerAttention. Because MatchGroupedAttention and MatchAttentionLayout are just matching between torch_attention_sdpa and torch_attention (the new unified grouped op), I can iterate through all the parameter combinations (64 patterns for MatchGroupedAttention and 16 for MatchAttentionLayout)
But eager attention's pattern is much more flexible, e.g. when/whether to cast softmax output can be a new feature to compose the pattern. Iterating through all these features can blow up the number of patterns quickly. In this case, I think it's okay to add new patterns manually when we discover a new way to write eager attention. Let me know your suggestions!
| "sinks": "sinks", | ||
| "sliding_window": "sliding_window", | ||
| "logit_cap": "logit_cap", | ||
| "layout": "bsnd", |
There was a problem hiding this comment.
this is arg name mapping between **kwargs names that huggingface may insert and **kwargs name we expect.
This doesn't make sense in this context
| torch.ops.auto_deploy.torch_attention_bsnd_grouped_sdpa, | ||
| torch.ops.auto_deploy.torch_attention, | ||
| args=(q_fake, k_fake, v_fake), | ||
| kwargs=node_kwargs, |
There was a problem hiding this comment.
Requires hard-coding layout="bsnd"
| def get_source_attention_op(cls) -> OpOverloadPacket: | ||
| """Get the source attention op that we target for replacement.""" | ||
| return torch.ops.auto_deploy.torch_attention_bsnd_grouped_sdpa | ||
| return torch.ops.auto_deploy.torch_attention |
There was a problem hiding this comment.
We need a sanity check now that torch_attention is called with layout "bsnd".
Probably best to put that into the get_constants utility and raise an error if there is a mismatch
There was a problem hiding this comment.
also needed for the other backends
| sinks: Optional[torch.Tensor] = None, | ||
| sliding_window: Optional[int] = None, | ||
| logit_cap: Optional[float] = None, | ||
| layout: str = "bnsd", # "bnsd" or "bsnd" |
There was a problem hiding this comment.
nit: you could type hint as
| layout: str = "bnsd", # "bnsd" or "bsnd" | |
| layout: Literal["bnsd", "bsnd"] = "bnsd", |
| for repeat_kv in (False, True): | ||
| for is_causal in (False, True): | ||
| for has_scale in (False, True): | ||
| for enable_gqa in (False, True): | ||
| for has_attn_mask in (False, True): | ||
| for has_dropout in (False, True): |
There was a problem hiding this comment.
good use case for itertools.product
| body_lines.append(call_line) | ||
| src = "\n".join(body_lines) | ||
| scope = {"torch": torch} | ||
| exec(src, scope) |
There was a problem hiding this comment.
no need for exec anywhere here. this can be solved with a "function factory", i.e., along the lines of
def make_multiplier(factor):
def multiply(x):
return x * factor
return multiply
Signed-off-by: Fridah-nv <201670829+Fridah-nv@users.noreply.github.com>
This PR does the following:
Summary by CodeRabbit
New Features
Refactor
Documentation
Tests
Description
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.