[TRTLLM-13948][feat] DeepSeek V3 Lite TranceiverV2 migration

[asfiyab-nvidia]

Summary by CodeRabbit

• Tests
  • Updated DeepSeek disaggregated serving configurations to use the NIXL cache-transfer backend.
  • Explicitly configured the Python transceiver runtime across context and generation server scenarios.
  • Applied the settings consistently across attention, overlap, CUDA graph, MTP, block reuse, and generation-only test variants.

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)

• If PR introduces API changes, an appropriate PR label is added - either api-compatible or api-breaking. For api-breaking, include BREAKING in the PR title.

• Any new dependencies have been scanned for license and vulnerabilities

• CODEOWNERS updated if ownership changes

• Documentation updated as needed

• Update tava architecture diagram if there is a significant design change in PR.

• 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

To see a list of available CI bot commands, please comment /bot help.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: d747f481-486d-43b3-ad59-adceda2c9571

📥 Commits

Reviewing files that changed from the base of the PR and between 6027183 and e8cf92d.

📒 Files selected for processing (18)

• tests/integration/defs/accuracy/test_disaggregated_serving.py
• tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp1_gentp1_deepseek_v3_lite.yaml
• tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp1_gentp1_deepseek_v3_lite_one_mtp.yaml
• tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp1_gentp1_deepseek_v3_lite_one_mtp_attention_dp_overlap.yaml
• tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp1_gentp1_deepseek_v3_lite_one_mtp_ctxpp2_gentp2.yaml
• tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp1_gentp1_deepseek_v3_lite_two_mtp.yaml
• tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp2_gentp2_deepseek_v3_lite.yaml
• tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp2_gentp2_deepseek_v3_lite_attention_dp.yaml
• tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp2_gentp2_deepseek_v3_lite_attention_dp_one.yaml
• tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp2_gentp2_deepseek_v3_lite_attention_dp_one_mtp.yaml
• tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp2_gentp2_deepseek_v3_lite_attention_dp_overlap.yaml
• tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp2_gentp2_deepseek_v3_lite_attention_dp_overlap_cuda_graph.yaml
• tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp2_gentp2_deepseek_v3_lite_nixl.yaml
• tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp2_gentp2_deepseek_v3_lite_overlap_cuda_graph.yaml
• tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp2ep2pp2_gentp4_deepseek_v3_lite_one_mtp_block_reuse.yaml
• tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp2ep2pp2_gentp4_deepseek_v3_lite_one_mtp_block_reuse_chunked.yaml
• tests/integration/defs/disaggregated/test_configs/disagg_config_deepseek_v3_lite_empty_batch.yaml
• tests/integration/defs/disaggregated/test_configs/disagg_config_gentp2_deepseek_v3_lite_attention_dp_gen_only.yaml

---

📝 Walkthrough

Walkthrough

The PR updates DeepSeek disaggregated serving configurations to use the NIXL cache backend with an explicit Python transceiver runtime, including the corresponding NIXL accuracy test configuration.

Changes

NIXL Python runtime configuration

Layer / File(s)	Summary
Disaggregated configuration updates tests/integration/defs/disaggregated/test_configs/*deepseek_v3_lite*.yaml	Cache transceiver settings across context and generation server configurations now use backend: NIXL and transceiver_runtime: PYTHON.
Accuracy test runtime alignment tests/integration/defs/accuracy/test_disaggregated_serving.py	The DeepSeek NIXL accuracy test explicitly sets the Python transceiver runtime for both server configurations.

Estimated code review effort: 2 (Simple) | ~10 minutes

Suggested reviewers: tabrizian

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Description check	⚠️ Warning	The template is present, but the Description and Test Coverage sections are empty, so the PR details are incomplete.	Add a short issue/solution summary and list the relevant tests or validation steps; keep the checklist if applicable.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title matches the main change and follows the required ticket/type format.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands.}

[asfiyab-nvidia]

/bot run --disable-fail-fast --stage-list "DGX_H100-4_GPUs-PyTorch-DeepSeek-1, DGX_B200-8_GPUs-PyTorch-1,
DGX_B200-8_GPUs-PyTorch-2, DGX_B200-8_GPUs-PyTorch-3, DGX_B200-8_GPUs-PyTorch-4"

[tensorrt-cicd]

PR_Github #59327 [ run ] triggered by Bot. Commit: d93a956 Link to invocation

[tensorrt-cicd]

PR_Github #59327 [ run ] completed with state SUCCESS. Commit: d93a956
/LLM/main/L0_MergeRequest_PR pipeline #47809 (Partly Tested) completed with status: 'SUCCESS'

CI Report

Link to invocation

[asfiyab-nvidia]

/bot run --disable-reuse-test --disable-fail-fast --stage-list "DGX_H100-4_GPUs-PyTorch-DeepSeek-1,
GB200-8_GPUs-2_Nodes-PyTorch-Disagg-PerfSanity-CTX1-NODE1-GPU1-GEN1-NODE1-GPU2-Post-Merge-1"

[tensorrt-cicd]

PR_Github #59489 [ run ] triggered by Bot. Commit: d93a956 Link to invocation

[tensorrt-cicd]

PR_Github #59489 [ run ] completed with state SUCCESS. Commit: d93a956
/LLM/main/L0_MergeRequest_PR pipeline #47949 (Partly Tested) completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

• Please check the failed tests and fix your PR
• If you cannot view the failures, ask the CI triggerer to share details
• Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation

[asfiyab-nvidia]

/bot run --disable-reuse-test --disable-fail-fast --stage-list "DGX_H100-4_GPUs-PyTorch-DeepSeek-1,
GB200-8_GPUs-2_Nodes-PyTorch-Disagg-PerfSanity-CTX1-NODE1-GPU1-GEN1-NODE1-GPU2-Post-Merge-1"

[tensorrt-cicd]

PR_Github #59531 [ run ] triggered by Bot. Commit: d93a956 Link to invocation

[tensorrt-cicd]

PR_Github #59531 [ run ] completed with state FAILURE. Commit: d93a956
/LLM/main/L0_MergeRequest_PR pipeline #47983 (Partly Tested) completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

• Please check the failed tests and fix your PR
• If you cannot view the failures, ask the CI triggerer to share details
• Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation

[yingguo-trt]

This config is shared by two tests, and hardcoding the backend here silently breaks one of them.

test_disaggregated_deepseek_v3_lite_fp8_ucx_tp1_single_gpu selects UCX by setting TRTLLM_USE_UCX_KVCACHE=1 and reuses this same config (config_map["deepseek_v3_lite_fp8_tp1"]). That env var is only consulted when the backend is DEFAULT — CacheTransceiverConfig._resolve_default_backend() returns early for any explicit backend:

if self.backend != "DEFAULT":
return self.backend, None
After this change the env var is ignored with no warning, so that test runs NIXL + Python instead of UCX. It still passes, and it is now byte-for-byte identical to test_disaggregated_deepseek_v3_lite_fp8_tp1_single_gpu. We'd lose UCX coverage at tp1/single-GPU (the case is enabled in l0_h100.yml:184) and burn a pre-merge slot on a duplicate.

Suggest adding a dedicated disagg_config_ctxtp1_gentp1_deepseek_v3_lite_ucx.yaml with backend: UCX, registering it as deepseek_v3_lite_fp8_tp1_ucx in config_map, and pointing the UCX test at it (the env var can then be dropped).