Scope DSv4 TRT non-MTP change to B300 only

B200 reverts to feat-deepseek_v4-9aa3715: the 2dd03e6 image OOMs on B200's
smaller HBM at conc-256 once SWA scratch reuse is disabled. Only B300 moves
to 2dd03e6 + TRTLLM_DSV4_ENABLE_SWA_SCRATCH_REUSE=0 in its launcher.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: Oseltamivir

.github/configs/nvidia-master.yaml

@@ -1801,7 +1801,7 @@ dsv4-fp4-b200-vllm-agentic:
       - { tp: 8, ep: 8, dp-attn: true, offloading: cpu,  conc-list: [64, 128, 256] }
 
 dsv4-fp4-b200-trt:
-  image: ghcr.io#semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-2dd03e6
+  image: ghcr.io#semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-9aa3715
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: b200-dsv4

benchmarks/single_node/fixed_seq_len/dsv4_fp4_b200_trt.sh

@@ -47,13 +47,6 @@ sanitize_slurm_mpi_env_for_trtllm
 export NCCL_NVLS_ENABLE="${NCCL_NVLS_ENABLE:-0}"
 echo "NCCL_NVLS_ENABLE: $NCCL_NVLS_ENABLE"
 
-# Disable DSv4 SWA scratch reuse: with attention-DP the V2 scheduler grows ctx KV
-# (incl. SWA scratch) before delay batching defers a request, and the resulting
-# revert_allocate_context -> resize(shrink) can't release the scratch of a
-# never-forwarded request (no finish_event), crashing every dpa=true job.
-export TRTLLM_DSV4_ENABLE_SWA_SCRATCH_REUSE="${TRTLLM_DSV4_ENABLE_SWA_SCRATCH_REUSE:-0}"
-echo "TRTLLM_DSV4_ENABLE_SWA_SCRATCH_REUSE: $TRTLLM_DSV4_ENABLE_SWA_SCRATCH_REUSE"
-
 if [[ "$MODEL" != /* ]]; then
     hf download "$MODEL"
 fi

perf-changelog.yaml

@@ -3387,10 +3387,9 @@
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1630
 
 - config-keys:
-    - dsv4-fp4-b200-trt
     - dsv4-fp4-b300-trt
   description:
-    - "Revert the non-MTP TensorRT-LLM DeepSeek-V4-Pro image to ghcr.io/semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-2dd03e6 and disable DSv4 SWA scratch reuse via TRTLLM_DSV4_ENABLE_SWA_SCRATCH_REUSE=0 in the launcher. Root cause of the prior attention-DP hangs/crashes: the V2 scheduler grows a context request's KV cache (incl. SWA scratch slots) before delay batching can defer it, so revert_allocate_context -> resize(shrink) must release scratch slots of a never-forwarded request, which has no finish_event -> crash on every dpa=true job. Disabling scratch reuse stops those slots from being allocated so the revert shrinks cleanly."
+    - "Update the non-MTP B300 TensorRT-LLM DeepSeek-V4-Pro image to ghcr.io/semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-2dd03e6 and disable DSv4 SWA scratch reuse via TRTLLM_DSV4_ENABLE_SWA_SCRATCH_REUSE=0 in the B300 launcher. Without this, attention-DP (dpa=true) jobs hang/crash: the V2 scheduler grows a context request's KV cache (incl. SWA scratch slots) before delay batching can defer it, so revert_allocate_context -> resize(shrink) must release scratch slots of a never-forwarded request, which has no finish_event -> crash on every dpa=true job. Disabling scratch reuse stops those slots from being allocated so the revert shrinks cleanly. B200 stays on feat-deepseek_v4-9aa3715 for now (the 2dd03e6 image OOMs on B200's smaller HBM at conc-256 with scratch reuse disabled)."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1636
 
 - config-keys: