From a0cecc2a085be8ea04feaf04ddc20372279f47a7 Mon Sep 17 00:00:00 2001 From: "claude[bot]" <41898282+claude[bot]@users.noreply.github.com> Date: Tue, 27 Oct 2026 16:18:15 +0000 Subject: [PATCH 1/7] perf(dsv4): update GB300 Dynamo-TRT AgentX image to TensorRT-LLM 1.4.0rc1 Co-Authored-By: Claude Opus 5.5 --- .../dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml | 2 +- inferencex-e2e/configs/nvidia-master.yaml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml index 7608b30e92..665ed19499 100644 --- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml +++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml @@ -6,7 +6,7 @@ base: model: {} container: {} frameworks: - tensorrt_llm: 1.3.0rc24 + tensorrt_llm: 1.4.0rc1 dynamo: install: true source: diff --git a/inferencex-e2e/configs/nvidia-master.yaml b/inferencex-e2e/configs/nvidia-master.yaml index 9ead101629..3c0de4738c 100644 --- a/inferencex-e2e/configs/nvidia-master.yaml +++ b/inferencex-e2e/configs/nvidia-master.yaml @@ -6504,7 +6504,7 @@ kimik3-fp4-gb200-dynamo-vllm-agentic-mooncake-dcp16-agg: dp-attn: false dsv4-fp4-gb300-dynamo-trt-agentx: - image: nvcr.io#nvidia/tensorrt-llm/release:1.3.0rc24 + image: nvcr.io/nvidia/tensorrt-llm/release:1.4.0rc1@sha256:3bedbe0321e7eb5430f3177d88894a3cf2dc2f6a087d0e85c13fc882667c75c0 model: deepseek-ai/DeepSeek-V4-Pro model-prefix: dsv4 runner: cluster:gb300-nv From 1fd1906bc58763850e55670b1287cf25924dc4d3 Mon Sep 17 00:00:00 2001 From: "claude[bot]" <41898282+claude[bot]@users.noreply.github.com> Date: Tue, 27 Oct 2026 16:18:46 +0000 Subject: [PATCH 2/7] chore(perf-changelog): record DSv4 GB300 Dynamo-TRT AgentX TensorRT-LLM 1.4.0rc1 update Co-Authored-By: Claude Opus 5.5 --- inferencex-e2e/perf-changelog.yaml | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml index fd894cc02f..df32fa7328 100644 --- a/inferencex-e2e/perf-changelog.yaml +++ b/inferencex-e2e/perf-changelog.yaml @@ -11413,3 +11413,10 @@ - Restore the MI355X MiniMax-M3 FP4 vLLM AgentX TP2/TP4 grid from run 34609443618 attempt 2 using SRT recipes. Update nightly-2a02f6efe319c885e3ccbcecde402e0028f9ec1e to nightly-df6603f0583375f244d1ac6d36265462b7490392 (2026-10-22, digest sha256:81899892d73777a4ae81e16f39e749169687f29249505eb4c56da83bd097d41c). - Enable upstream MiniMax-M3 monokernel at TP4 with Model Runner V2, AITER MoE, FP8 KV/indexer and block size 128; keep TP2 on ordinary decode. Preserve EAGLE3-GQA with three draft tokens and use automatic golden acceptance for throughput with real verification for evals. pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4418 + +- config-keys: + - dsv4-fp4-gb300-dynamo-trt-agentx + description: + - "Update the TensorRT-LLM image from 1.3.0rc24 to 1.4.0rc1, pinned by digest." + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4583 + scenario-type: [agentic-coding] From df494afcf66a5ece9c6292c5fc8039d44da6d9f1 Mon Sep 17 00:00:00 2001 From: "claude[bot]" <41898282+claude[bot]@users.noreply.github.com> Date: Tue, 27 Oct 2026 16:45:30 +0000 Subject: [PATCH 3/7] fix(dsv4): omit --publish-metrics for the pinned Dynamo fork on GB300 Dynamo-TRT AgentX Co-Authored-By: Claude Opus 5.5 --- .../dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml | 2 ++ inferencex-e2e/perf-changelog.yaml | 2 +- 2 files changed, 3 insertions(+), 1 deletion(-) diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml index 665ed19499..bc929c8e0d 100644 --- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml +++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml @@ -22,6 +22,8 @@ base: engine: type: trtllm publish_events_and_metrics: false + # Dynamo builds before ai-dynamo/dynamo#12162 reject --publish-metrics. + publish_metrics: false roles: prefill: gpus: 8 diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml index df32fa7328..dc5f380e51 100644 --- a/inferencex-e2e/perf-changelog.yaml +++ b/inferencex-e2e/perf-changelog.yaml @@ -11417,6 +11417,6 @@ - config-keys: - dsv4-fp4-gb300-dynamo-trt-agentx description: - - "Update the TensorRT-LLM image from 1.3.0rc24 to 1.4.0rc1, pinned by digest." + - "Update the TensorRT-LLM image from 1.3.0rc24 to 1.4.0rc1 and omit the --publish-metrics Dynamo flag." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4583 scenario-type: [agentic-coding] From d99b87480acae5ea40ba605bed7eb32f32bc7abe Mon Sep 17 00:00:00 2001 From: "claude[bot]" <41898282+claude[bot]@users.noreply.github.com> Date: Tue, 27 Oct 2026 20:27:25 +0000 Subject: [PATCH 4/7] fix(dsv4): pin upstream Dynamo v1.6.1 for GB300 Dynamo-TRT AgentX The Dynamo fork pin (upstream 2026-08-07) cannot parse an interface-name DYN_TCP_RPC_HOST, which the batch3 GB300 nodes export; upstream resolves it since ai-dynamo/dynamo#11870 and v1.6.1 also contains #12162. Co-Authored-By: Claude Opus 5.5 --- .../dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml | 8 +++++--- inferencex-e2e/perf-changelog.yaml | 2 +- 2 files changed, 6 insertions(+), 4 deletions(-) diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml index bc929c8e0d..d4ae4a85b0 100644 --- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml +++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml @@ -10,8 +10,10 @@ base: dynamo: install: true source: - git: https://github.com/cquil11/dynamo.git - rev: 2cbbdc863c336c2cfffa2a3668e2a6661801b259 + # Upstream v1.6.1: resolves interface-name DYN_TCP_RPC_HOST (ai-dynamo/dynamo#11870) + # and decouples TRT-LLM KV events from metrics (ai-dynamo/dynamo#12162). + git: https://github.com/ai-dynamo/dynamo.git + rev: 0d9c24a8bf6db13c1a1223cf15fcbee9140d1ac3 request_plane: tcp health_check: max_attempts: 1440 @@ -22,7 +24,7 @@ base: engine: type: trtllm publish_events_and_metrics: false - # Dynamo builds before ai-dynamo/dynamo#12162 reject --publish-metrics. + # Keep metrics publishing off, as with the earlier Dynamo pin. publish_metrics: false roles: prefill: diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml index 4a2473f2c5..833aa7d54a 100644 --- a/inferencex-e2e/perf-changelog.yaml +++ b/inferencex-e2e/perf-changelog.yaml @@ -11433,6 +11433,6 @@ - config-keys: - dsv4-fp4-gb300-dynamo-trt-agentx description: - - "Update the TensorRT-LLM image from 1.3.0rc24 to 1.4.0rc1 and omit the --publish-metrics Dynamo flag." + - "Update the TensorRT-LLM image from 1.3.0rc24 to 1.4.0rc1 and the recipe Dynamo pin to v1.6.1." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4583 scenario-type: [agentic-coding] From 7abc13f3fb3c6b5abfefa0173fe5edf29548ff86 Mon Sep 17 00:00:00 2001 From: "claude[bot]" <41898282+claude[bot]@users.noreply.github.com> Date: Tue, 27 Oct 2026 20:50:53 +0000 Subject: [PATCH 5/7] fix(dsv4): install Dynamo v1.6.1 from PyPI wheels for GB300 Dynamo-TRT AgentX The v1.6.1 source build fails in the TensorRT-LLM 1.4.0rc1 image because cudarc 0.19.8 rejects CUDA toolkit 13.4; the PyPI wheels ship the same release. Co-Authored-By: Claude Opus 5.5 --- .../dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml index d4ae4a85b0..4d76adf227 100644 --- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml +++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml @@ -12,8 +12,8 @@ base: source: # Upstream v1.6.1: resolves interface-name DYN_TCP_RPC_HOST (ai-dynamo/dynamo#11870) # and decouples TRT-LLM KV events from metrics (ai-dynamo/dynamo#12162). - git: https://github.com/ai-dynamo/dynamo.git - rev: 0d9c24a8bf6db13c1a1223cf15fcbee9140d1ac3 + # Prebuilt wheels: a source build's cudarc 0.19.8 rejects the image's CUDA 13.4. + pypi: "1.6.1" request_plane: tcp health_check: max_attempts: 1440 From 68cb251dd958f8e1a9377c8d60634f51bc96fa0d Mon Sep 17 00:00:00 2001 From: "claude[bot]" <41898282+claude[bot]@users.noreply.github.com> Date: Tue, 27 Oct 2026 22:56:18 +0000 Subject: [PATCH 6/7] fix(dsv4): disable SWA scratch reuse on GB300 Dynamo-TRT AgentX prefill for TensorRT-LLM 1.4.0rc1 Co-Authored-By: Claude Opus 5.5 --- .../dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml | 3 +++ 1 file changed, 3 insertions(+) diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml index 4d76adf227..f53207426e 100644 --- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml +++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml @@ -60,6 +60,9 @@ base: event_buffer_max_size: 0 free_gpu_memory_fraction: 0.8 host_cache_size: 193273528320 + # 1.4.0rc1 always enforces the SWA scratch resize check (debug-only in 1.3.0rc24), + # which c388 prefill tripped; DeepSeek-V4 model defaults turn scratch reuse on. + enable_swa_scratch_reuse: false pool_ratio: - 0.55 - 0.22 From 66b9743f66ce5676b5e356404902e41b739cab0c Mon Sep 17 00:00:00 2001 From: "claude[bot]" <41898282+claude[bot]@users.noreply.github.com> Date: Wed, 28 Oct 2026 01:58:20 +0000 Subject: [PATCH 7/7] fix(dsv4): publish TRT-LLM metrics on GB300 Dynamo-TRT AgentX with Dynamo 1.6.1 Dynamo 1.6.1 accepts --publish-metrics (ai-dynamo/dynamo#12162), and the recipe requires trtllm_kv_cache_utilization in the frontend metrics export. Co-Authored-By: Claude Opus 5.5 --- .../dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml | 2 -- 1 file changed, 2 deletions(-) diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml index f53207426e..6d01e494bf 100644 --- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml +++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml @@ -24,8 +24,6 @@ base: engine: type: trtllm publish_events_and_metrics: false - # Keep metrics publishing off, as with the earlier Dynamo pin. - publish_metrics: false roles: prefill: gpus: 8