[https://nvbugs/5271281][fix] fix a pd+mtp accuracy issue

[lfr-0531]

Description

There are two issues:

• In context server, the max_new_tokens=1. In mtp sampler, we will get True when calling _handle_stop_criteria, and the request.py_draft_tokens won't be updated. So in the first iteration of the generation requests, the draft tokens are not correct. Fixed it in mtp sampler.
• When transferring parameters from context server to the generation server, the schedued generation requests will be updated after preparing the resouces, including the draft tokens. As a result, before the first generation forward, we didn't add new space in kv cache manager for those draft tokens. Fixed it in py_executor.

Test Coverage

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".

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.

[lfr-0531]

/bot run

[tensorrt-cicd]

PR_Github #6011 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #6011 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #4392 completed with status: 'FAILURE'

[lfr-0531]

/bot run

[tensorrt-cicd]

PR_Github #6044 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #6044 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #4416 completed with status: 'SUCCESS'

[Shixiaowei02]

Please get the approval from the NVIDIA/trt-llm-torch-devs group. @lfr-0531

[pcastonguay]

We should have a test that verifies this fixes the accuracy issue. We have some accuracy tests for DP, can we add coverage for PD+MTP?

[lfr-0531]

We should have a test that verifies this fixes the accuracy issue. We have some accuracy tests for DP, can we add coverage for PD+MTP?

Hi @pcastonguay, do you mean accuracy tests on some datasets? or just compare the output texts? I saw all of the disaggregated servering tests compare the output texts. And we already have the test for pd + mtp + attention dp + overlap scheduler.

TensorRT-LLM/tests/integration/defs/disaggregated/test_disaggregated.py

Line 104 in 419151f

"deepseek_v3_lite_fp8_tp1_attention_dp_overlap_one_mtp":

If the former one, it is not an easy work to cover PD cases in our accuracy tests. It would be better to add it in a new PR.

[pcastonguay]

We should have a test that verifies this fixes the accuracy issue. We have some accuracy tests for DP, can we add coverage for PD+MTP?

Hi @pcastonguay, do you mean accuracy tests on some datasets? or just compare the output texts? I saw all of the disaggregated servering tests compare the output texts. And we already have the test for pd + mtp + attention dp + overlap scheduler.

TensorRT-LLM/tests/integration/defs/disaggregated/test_disaggregated.py

Line 104 in 419151f

"deepseek_v3_lite_fp8_tp1_attention_dp_overlap_one_mtp":

If the former one, it is not an easy work to cover PD cases in our accuracy tests. It would be better to add it in a new PR.

The former. We already have accuracy tests with disagg: https://github.com/NVIDIA/TensorRT-LLM/blob/main/tests/integration/defs/accuracy/test_disaggregated_serving.py. Could we extend to cover PD + MTP + Attention DP + overlap? @Tabrizian for vis.

[Tabrizian]

@lfr-0531 You can specify the pytorch config needed for your setup here:

TensorRT-LLM/tests/integration/defs/accuracy/test_disaggregated_serving.py

Lines 181 to 190 in ca3eaf4

	ctx_server_config = {
	"pytorch_backend_config": {
	"disable_overlap_scheduler": True
	}
	}
	gen_server_config = {
	"pytorch_backend_config": {
	"disable_overlap_scheduler": disable_overlap_scheduler
	}
	}

[lfr-0531]

Thanks, @pcastonguay and @Tabrizian! I added the pd+mtp accuracy test.

[lfr-0531]

/bot run

[tensorrt-cicd]

PR_Github #6486 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #6486 [ run ] completed with state FAILURE
/LLM/main/L0_MergeRequest_PR pipeline #4746 completed with status: 'FAILURE'

[lfr-0531]

/bot run

[tensorrt-cicd]

PR_Github #6489 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #6489 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #4749 completed with status: 'FAILURE'

[Tabrizian]

@lfr-0531 It looks like the new test that was added was never triggered in any of these pipelines. I believe you needed to trigger it manually.

[lfr-0531]

@lfr-0531 It looks like the new test that was added was never triggered in any of these pipelines. I believe you needed to trigger it manually.

Ah, I added them to the post merge. I saw that those tests are waived in nvbugs/5322354. Let's wait for the mass integration. After that, I'll rerun the tests.