From 860d76e311fc927a1510b029d1ecfa83e88cccc3 Mon Sep 17 00:00:00 2001 From: Rohit Pujar Nagraj Date: Tue, 8 Sep 2026 10:02:50 -0700 Subject: [PATCH 1/3] chore(dsv4): rerun GB300 TRT-LLM AgentX MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 重新运行 GB300 Dynamo TensorRT-LLM AgentX 配置，以验证 KV 卸载环境变量的传递。 --- perf-changelog.yaml | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 4fcd229bfb..d38a9c344d 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -6958,3 +6958,10 @@ - "Expand the TP8 and TP4/EP1 sweep coverage through concurrency 640." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2866 +- config-keys: + - dsv4-fp4-gb300-dynamo-trt-agentx + scenario-type: + - agentic-coding + description: + - "Re-run the GB300 Dynamo TensorRT-LLM AgentX configuration to verify KV offload environment propagation." + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX From 81f683b7e4c4ca3948e4de45900fb159613e1f3d Mon Sep 17 00:00:00 2001 From: Rohit Pujar Nagraj Date: Tue, 8 Sep 2026 10:03:41 -0700 Subject: [PATCH 2/3] chore(changelog): link PR #2891 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 在性能变更日志中补充 PR #2891 链接。 --- perf-changelog.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index d38a9c344d..36263c3963 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -5595,7 +5595,7 @@ - "Image ghcr.io/tile-ai/tilert:0.1.5 (tilert 0.1.5.post2 installed at container start); commands aligned to TileRT README Topology A -- NIXL KV transfer, --kv-cache-dtype fp8_ds_mla (prefill) <-> fp8 (decode), max-seq-len 202752; MTP speculative-config wired via spec-decoding=mtp" - "Topology: 1 prefill node (TP8) + 1 decode node (TP8), each 8xB200 exclusive; TileRT decode is bs=1 only so conc-list is a single point [1], ISL 1k/8k OSL 1k" - "Runner: launch_b200-dgxc.sh tilert early-return branch (zero impact on the dynamo path); tilert_utils/submit.sh issues two srun --ntasks=1, one per role, because prefill and decode need different container images; roles are dispatched by the TILERT_ROLE it exports, and torn down across nodes via a sentinel file on the shared /workspace" - pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2891 - config-keys: - qwen3.5-fp8-b200-sglang-agentic-mtp From 7dcd29b2f88d50b6354280e37bdb3084bba274d7 Mon Sep 17 00:00:00 2001 From: Rohit Pujar Nagraj Date: Tue, 8 Sep 2026 10:04:57 -0700 Subject: [PATCH 3/3] fix(changelog): correct PR #2891 link MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 恢复既有变更日志条目，并将 PR #2891 链接写入本次重新运行条目。 --- perf-changelog.yaml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 36263c3963..7668d8d771 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -5595,7 +5595,7 @@ - "Image ghcr.io/tile-ai/tilert:0.1.5 (tilert 0.1.5.post2 installed at container start); commands aligned to TileRT README Topology A -- NIXL KV transfer, --kv-cache-dtype fp8_ds_mla (prefill) <-> fp8 (decode), max-seq-len 202752; MTP speculative-config wired via spec-decoding=mtp" - "Topology: 1 prefill node (TP8) + 1 decode node (TP8), each 8xB200 exclusive; TileRT decode is bs=1 only so conc-list is a single point [1], ISL 1k/8k OSL 1k" - "Runner: launch_b200-dgxc.sh tilert early-return branch (zero impact on the dynamo path); tilert_utils/submit.sh issues two srun --ntasks=1, one per role, because prefill and decode need different container images; roles are dispatched by the TILERT_ROLE it exports, and torn down across nodes via a sentinel file on the shared /workspace" - pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2891 + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX - config-keys: - qwen3.5-fp8-b200-sglang-agentic-mtp @@ -6964,4 +6964,4 @@ - agentic-coding description: - "Re-run the GB300 Dynamo TensorRT-LLM AgentX configuration to verify KV offload environment propagation." - pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2891