[nvbug 5304752][fix] enhance _check_arguments to filter illegal requests for pytorch backend #5541

LinPoly · 2025-06-27T07:47:41Z

Description

This PR add check against the prompt length, to prevent a single illegal request from killing by the pytorch executor.

Test Coverage

Extend the backend of test_llm_capture_request_error to test pytorch backend.

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 [--disable-fail-fast --skip-test --stage-list "A10-1, xxx" --gpu-type "A30, H100_PCIe" --add-multi-gpu-test --only-multi-gpu-test --disable-multi-gpu-test --post-merge --extra-stage "H100_PCIe-[Post-Merge]-1, xxx"]

Launch build/test pipelines. All previously running jobs will be killed.

--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 "A10-1, xxx" (OPTIONAL) : Only run the specified test stages. Examples: "A10-1, 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.

--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. Will also run 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-[Post-Merge]-1, xxx" (OPTIONAL) : Run the ordinary L0 pre-merge pipeline and specified test stages. Examples: --extra-stage "H100_PCIe-[Post-Merge]-1, xxx".

For guidance on mapping tests to stage names, see docs/source/reference/ci-overview.md.

kill

kill

Kill all running builds associated with pull request.

skip

skip --comment COMMENT

Skip 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-pipeline

Reuse 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.

LinPoly · 2025-06-27T07:52:31Z

/bot run

tensorrt-cicd · 2025-06-27T07:57:55Z

PR_Github #10126 [ run ] triggered by Bot

tensorrt-cicd · 2025-06-27T09:58:07Z

PR_Github #10126 [ run ] completed with state SUCCESS
/LLM/release-0.21/L0_MergeRequest_PR pipeline #56 completed with status: 'FAILURE'

LinPoly · 2025-06-27T11:18:36Z

/bot run

tensorrt-cicd · 2025-06-27T11:23:52Z

PR_Github #10147 [ run ] triggered by Bot

syuoni

LGTM

tensorrt-cicd · 2025-06-27T17:02:05Z

PR_Github #10147 [ run ] completed with state SUCCESS
/LLM/release-0.21/L0_MergeRequest_PR pipeline #58 completed with status: 'SUCCESS'
Pipeline passed with automatic retried tests. Check the rerun report for details.

kaiyux · 2025-07-01T13:12:53Z

tensorrt_llm/llmapi/llm.py

@@ -529,8 +529,13 @@ def _check_arguments(self, prompt_len: int, query_len: int,
                raise ValueError(
                    f"PyTorch backend currently only supports `logprobs=1`. Received `logprobs={sampling_params.logprobs}` (Top{sampling_params.logprobs} logprobs). Please set `logprobs=1` in `sampling_params` instead."
                )
-            return
-        elif self.args.backend == "_autodeploy":


Is this line expected to be removed?

I think this line will not be touched, if backend == "_autodeploy" , execution will fall in the if statement in line 521, then it will return or raise error.

@lucaslie Can you help confirm if the changes are ok? Thanks.

Yes, that looks good. it's actually already removed on the main branch:

TensorRT-LLM/tensorrt_llm/llmapi/llm.py

Lines 522 to 533 in 2b0c87e

if self.args.backend in ["pytorch", "_autodeploy"]:

# TODO: remove these checks after PyTorch backend

# fully support TopK prompt and generation logprobs.

if sampling_params.prompt_logprobs:

raise ValueError(

f"`prompt_logprobs` in sampling_params is not supported in the PyTorch backend yet. Received `prompt_logprobs={sampling_params.prompt_logprobs}`. Please unset this field."

)

if sampling_params.logprobs and sampling_params.logprobs > 1:

raise ValueError(

f"PyTorch backend currently only supports `logprobs=1`. Received `logprobs={sampling_params.logprobs}` (Top{sampling_params.logprobs} logprobs). Please set `logprobs=1` in `sampling_params` instead."

)

return

kaiyux · 2025-07-01T13:13:49Z

tensorrt_llm/llmapi/llm.py

+            if not self.args.enable_chunked_prefill:
+                max_num_tokens = self.args.max_num_tokens
+                if max_num_tokens and prompt_len / self.args.parallel_config.cp_size + query_len > max_num_tokens:
+                    raise ValueError(


Ideally we should just return an error request instead of stop the system?

This will be catch by OpenaiServer and it will return error response to the client, the service is still alive

auto_deploy may not support the same features, so better to keep the skipping logic here.

That would be easy, another backend checking before this error raising is enough. Thanks and will do.

Superjomn · 2025-07-01T13:21:29Z

tensorrt_llm/llmapi/llm.py

+            if not self.args.enable_chunked_prefill:
+                max_num_tokens = self.args.max_num_tokens
+                if max_num_tokens and prompt_len / self.args.parallel_config.cp_size + query_len > max_num_tokens:
+                    raise ValueError(


auto_deploy may not support the same features, so better to keep the skipping logic here.

Superjomn · 2025-07-01T13:23:03Z

tests/unittest/llmapi/test_llm.py

    build_config = BuildConfig()
    build_config.max_num_tokens = 64

    llm = LLM(
        model=llama_model_path,
        build_config=build_config,
+        tensor_parallel_size=tp_size,


We are going to remove the backend argument, so it is better to from tensorrt_llm import LLM as TorchLLM and from tensorrt_llm._tensorrt_engine import LLM as TrtLLM and test with both.

Signed-off-by: Pengyun Lin <81065165+LinPoly@users.noreply.github.com>

LinPoly · 2025-07-02T07:03:26Z

/bot run

tensorrt-cicd · 2025-07-02T07:10:13Z

PR_Github #10584 [ run ] triggered by Bot

tensorrt-cicd · 2025-07-02T12:59:17Z

PR_Github #10584 [ run ] completed with state SUCCESS
/LLM/release-0.21/L0_MergeRequest_PR pipeline #106 completed with status: 'ABORTED'

LinPoly requested a review from a team as a code owner June 27, 2025 07:47

LinPoly requested a review from QiJune June 27, 2025 07:48

LinPoly changed the title ~~[nvbug 5304752][fix]: enhance _check_arguments to filter illegal requests~~ [nvbug 5304752][fix]: enhance _check_arguments to filter illegal requests for pytorch backend Jun 27, 2025

LinPoly requested a review from syuoni June 27, 2025 08:08

LinPoly force-pushed the user/pengyunl/enhance_check_args branch from d0023fa to 7635b98 Compare June 27, 2025 11:17

syuoni approved these changes Jun 27, 2025

View reviewed changes

LinPoly requested a review from kaiyux July 1, 2025 13:07

kaiyux reviewed Jul 1, 2025

View reviewed changes

kaiyux requested a review from Superjomn July 1, 2025 13:14

Superjomn reviewed Jul 1, 2025

View reviewed changes

LinPoly added 4 commits July 2, 2025 05:45

Remove untouched branch

78d1d3c

Signed-off-by: Pengyun Lin <81065165+LinPoly@users.noreply.github.com>

Enhance _check_arguments to filter illegal requests

ae5248d

Signed-off-by: Pengyun Lin <81065165+LinPoly@users.noreply.github.com>

Fix chunk prefill failure

41eece1

Signed-off-by: Pengyun Lin <81065165+LinPoly@users.noreply.github.com>

Refactor test

a036283

Signed-off-by: Pengyun Lin <81065165+LinPoly@users.noreply.github.com>

LinPoly force-pushed the user/pengyunl/enhance_check_args branch from 7635b98 to a036283 Compare July 2, 2025 07:03

LinPoly changed the title ~~[nvbug 5304752][fix]: enhance _check_arguments to filter illegal requests for pytorch backend~~ [nvbug 5304752][fix] enhance _check_arguments to filter illegal requests for pytorch backend Jul 4, 2025

	if self.args.backend in ["pytorch", "_autodeploy"]:
	# TODO: remove these checks after PyTorch backend
	# fully support TopK prompt and generation logprobs.
	if sampling_params.prompt_logprobs:
	raise ValueError(
	f"`prompt_logprobs` in sampling_params is not supported in the PyTorch backend yet. Received `prompt_logprobs={sampling_params.prompt_logprobs}`. Please unset this field."
	)
	if sampling_params.logprobs and sampling_params.logprobs > 1:
	raise ValueError(
	f"PyTorch backend currently only supports `logprobs=1`. Received `logprobs={sampling_params.logprobs}` (Top{sampling_params.logprobs} logprobs). Please set `logprobs=1` in `sampling_params` instead."
	)
	return

[nvbug 5304752][fix] enhance _check_arguments to filter illegal requests for pytorch backend #5541

Are you sure you want to change the base?

[nvbug 5304752][fix] enhance _check_arguments to filter illegal requests for pytorch backend #5541

Uh oh!

Conversation

LinPoly commented Jun 27, 2025

Description

Test Coverage

GitHub Bot Help

kill

skip

reuse-pipeline

Uh oh!

LinPoly commented Jun 27, 2025

Uh oh!

tensorrt-cicd commented Jun 27, 2025

Uh oh!

tensorrt-cicd commented Jun 27, 2025

Uh oh!

LinPoly commented Jun 27, 2025

Uh oh!

tensorrt-cicd commented Jun 27, 2025

Uh oh!

syuoni left a comment

Choose a reason for hiding this comment

Uh oh!

tensorrt-cicd commented Jun 27, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

lucaslie Jul 3, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

LinPoly commented Jul 2, 2025

Uh oh!

tensorrt-cicd commented Jul 2, 2025

Uh oh!

tensorrt-cicd commented Jul 2, 2025

Uh oh!

Uh oh!

lucaslie Jul 3, 2025 •

edited

Loading