From 4bcefb7da5353a24ea2e37cfe17de1195210c865 Mon Sep 17 00:00:00 2001
From: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Date: Mon, 1 Jun 2026 14:26:45 -0700
Subject: [PATCH 1/9] Update DSv4 TRT image for B200/B300 (non-MTP) to
 feat-deepseek_v4-2dd03e6

Bumps the TensorRT-LLM DeepSeek-V4-Pro image for dsv4-fp4-b200-trt and
dsv4-fp4-b300-trt to ghcr.io#semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-2dd03e6.
The -mtp variants are intentionally left on feat-deepseek_v4-9aa3715.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---
 .github/configs/nvidia-master.yaml | 4 ++--
 perf-changelog.yaml                | 7 +++++++
 2 files changed, 9 insertions(+), 2 deletions(-)

diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index baa0b38a1..887f2fc0b 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.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_v4-9aa3715
+  image: ghcr.io#semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-2dd03e6
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: b200-dsv4
@@ -3053,7 +3053,7 @@ dsv4-fp4-b300-vllm-agentic:
       - { tp: 8, ep: 8, dp-attn: true, offloading: cpu,  conc-list: [128, 256, 512] }
 
 dsv4-fp4-b300-trt:
-  image: ghcr.io#semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-9aa3715
+  image: ghcr.io#semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-2dd03e6
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: b300
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 263da4cfa..cb59c07eb 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -3354,3 +3354,10 @@
   description:
     - "Add MTP speculative-decoding sibling for dsv4-fp4-mi355x-vllm (model: deepseek-ai/DeepSeek-V4-Pro) on vllm/vllm-openai-rocm:v0.22.0, per vllm-project/vllm#43385"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1630
+
+- config-keys:
+    - dsv4-fp4-b200-trt
+    - dsv4-fp4-b300-trt
+  description:
+    - "Update the TensorRT-LLM DeepSeek-V4-Pro image to ghcr.io/semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-2dd03e6"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX

From 6b7558c481d00bdb0d005ccf481347ed8aff2a57 Mon Sep 17 00:00:00 2001
From: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Date: Mon, 1 Jun 2026 14:27:13 -0700
Subject: [PATCH 2/9] Backfill PR number in changelog pr-link

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index cb59c07eb..28932a990 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -3360,4 +3360,4 @@
     - dsv4-fp4-b300-trt
   description:
     - "Update the TensorRT-LLM DeepSeek-V4-Pro image to ghcr.io/semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-2dd03e6"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1636

From f441f9fbeb857f2c3cfd5066e83c71abeeef25e7 Mon Sep 17 00:00:00 2001
From: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Date: Mon, 1 Jun 2026 15:52:35 -0700
Subject: [PATCH 3/9] Try official TRT-LLM release image 1.3.0rc15.post1 for
 DSv4 B200/B300 (non-MTP)

Swap dsv4-fp4-b200-trt and dsv4-fp4-b300-trt from the custom
ghcr.io semianalysis feat/deepseek_v4 build to the official
nvcr.io#nvidia/tensorrt-llm/release:1.3.0rc15.post1 to test whether the
official RC can serve DeepSeek-V4-Pro. The -mtp variants are unchanged.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---
 .github/configs/nvidia-master.yaml | 4 ++--
 perf-changelog.yaml                | 2 +-
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index 7f1ac486d..a3bfcc5ae 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.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: nvcr.io#nvidia/tensorrt-llm/release:1.3.0rc15.post1
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: b200-dsv4
@@ -3049,7 +3049,7 @@ dsv4-fp4-b300-vllm-agentic:
       - { tp: 8, ep: 8, dp-attn: true, offloading: cpu,  conc-list: [128, 256, 512] }
 
 dsv4-fp4-b300-trt:
-  image: ghcr.io#semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-2dd03e6
+  image: nvcr.io#nvidia/tensorrt-llm/release:1.3.0rc15.post1
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: b300
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 8da6e5738..16ef3702f 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -3372,5 +3372,5 @@
     - dsv4-fp4-b200-trt
     - dsv4-fp4-b300-trt
   description:
-    - "Update the TensorRT-LLM DeepSeek-V4-Pro image to ghcr.io/semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-2dd03e6"
+    - "Try the official NVIDIA TensorRT-LLM release image nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc15.post1 for DeepSeek-V4-Pro (replacing the custom feat/deepseek_v4 build) to see whether the official RC can serve DSv4-Pro"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1636

From 4bc5592f8ef18c2ba8952c8f3b9a9020d439a225 Mon Sep 17 00:00:00 2001
From: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Date: Mon, 1 Jun 2026 23:47:28 -0700
Subject: [PATCH 4/9] Revert to custom feat/deepseek_v4-2dd03e6 image for DSv4
 B200/B300 (non-MTP)

The official nvcr.io tensorrt-llm/release:1.3.0rc15.post1 loads DSv4-Pro but
its DP-attention path deadlocks/crashes under concurrent load (every dpa=true
job hung or failed; only pure-TP conc-1 points passed). Revert to the stable
custom build until upstream fixes DSv4 + attention-DP (NVIDIA/TensorRT-LLM#13431).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---
 .github/configs/nvidia-master.yaml | 4 ++--
 perf-changelog.yaml                | 2 +-
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index a3bfcc5ae..7f1ac486d 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.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: nvcr.io#nvidia/tensorrt-llm/release:1.3.0rc15.post1
+  image: ghcr.io#semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-2dd03e6
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: b200-dsv4
@@ -3049,7 +3049,7 @@ dsv4-fp4-b300-vllm-agentic:
       - { tp: 8, ep: 8, dp-attn: true, offloading: cpu,  conc-list: [128, 256, 512] }
 
 dsv4-fp4-b300-trt:
-  image: nvcr.io#nvidia/tensorrt-llm/release:1.3.0rc15.post1
+  image: ghcr.io#semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-2dd03e6
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: b300
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 16ef3702f..8da6e5738 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -3372,5 +3372,5 @@
     - dsv4-fp4-b200-trt
     - dsv4-fp4-b300-trt
   description:
-    - "Try the official NVIDIA TensorRT-LLM release image nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc15.post1 for DeepSeek-V4-Pro (replacing the custom feat/deepseek_v4 build) to see whether the official RC can serve DSv4-Pro"
+    - "Update the TensorRT-LLM DeepSeek-V4-Pro image to ghcr.io/semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-2dd03e6"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1636

From 14a1bb32713c1b24741a987d3460bce6a45d7cda Mon Sep 17 00:00:00 2001
From: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Date: Tue, 2 Jun 2026 02:39:16 -0700
Subject: [PATCH 5/9] Point DSv4 B200/B300 TRT (non-MTP) at the SWA-scratch-fix
 image

Bump dsv4-fp4-b200-trt and dsv4-fp4-b300-trt to
ghcr.io#semianalysisai/trtllm-deepseek-v4:fix-dsv4-swa-scratch-revert-shrink-c914d6d
(TRT-LLM feat/deepseek_v4 @ 084cf2ba + kv_cache_manager_v2 fix). This resolves
the engine crash on attention-DP context/generation reverts at high concurrency
(the b300 8k1k conc>=512 "LLM is shutting down" hang). The -mtp variants stay on
feat-deepseek_v4-9aa3715.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---
 .github/configs/nvidia-master.yaml | 4 ++--
 perf-changelog.yaml                | 2 +-
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index 87ebd7c2e..8dfac57b5 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.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:fix-dsv4-swa-scratch-revert-shrink-c914d6d
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: b200-dsv4
@@ -3049,7 +3049,7 @@ dsv4-fp4-b300-vllm-agentic:
       - { tp: 8, ep: 8, dp-attn: true, offloading: cpu,  conc-list: [128, 256, 512] }
 
 dsv4-fp4-b300-trt:
-  image: ghcr.io#semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-2dd03e6
+  image: ghcr.io#semianalysisai/trtllm-deepseek-v4:fix-dsv4-swa-scratch-revert-shrink-c914d6d
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: b300
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index ddd6c84fc..ae99f0455 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -3372,7 +3372,7 @@
     - dsv4-fp4-b200-trt
     - dsv4-fp4-b300-trt
   description:
-    - "Update the TensorRT-LLM DeepSeek-V4-Pro image to ghcr.io/semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-2dd03e6"
+    - "Update the TensorRT-LLM DeepSeek-V4-Pro image to ghcr.io/semianalysisai/trtllm-deepseek-v4:fix-dsv4-swa-scratch-revert-shrink-c914d6d (TRT-LLM feat/deepseek_v4 @ 084cf2ba + kv_cache_manager_v2 fix: free SWA scratch slots on shrink instead of asserting, which crashed the engine on attention-DP context/generation reverts at high concurrency, e.g. b300 8k1k conc>=512)"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1636
 
 - config-keys:

From e23a5415c387e4c81bde634aaa926d8fdfdc0730 Mon Sep 17 00:00:00 2001
From: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Date: Tue, 2 Jun 2026 12:32:38 -0700
Subject: [PATCH 6/9] Revert DSv4 B200/B300 TRT (non-MTP) to 2dd03e6 + disable
 SWA scratch reuse

The c914d6d image's kv_cache_manager_v2 patch was wrong: freeing SWA scratch
slots on the attention-DP revert->resize(shrink) path hits finish_event=None
(a deferred request never forwarded), crashing every dpa=true job and hanging
the engine. Root cause is a V2-scheduler / SWA-scratch-reuse conflict: the V2
scheduler grows a context request's KV cache (incl. SWA scratch) before delay
batching can defer it, so revert_allocate_context -> resize(shrink) must release
scratch slots that have no finish_event.

Revert both non-MTP images to feat-deepseek_v4-2dd03e6 and set
TRTLLM_DSV4_ENABLE_SWA_SCRATCH_REUSE=0 in the launchers so no scratch slots are
allocated and the revert shrinks cleanly. MTP configs untouched (9aa3715).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---
 .github/configs/nvidia-master.yaml                        | 4 ++--
 benchmarks/single_node/fixed_seq_len/dsv4_fp4_b200_trt.sh | 7 +++++++
 benchmarks/single_node/fixed_seq_len/dsv4_fp4_b300_trt.sh | 7 +++++++
 perf-changelog.yaml                                       | 2 +-
 4 files changed, 17 insertions(+), 3 deletions(-)

diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index f3837a606..718f95196 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.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:fix-dsv4-swa-scratch-revert-shrink-c914d6d
+  image: ghcr.io#semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-2dd03e6
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: b200-dsv4
@@ -3049,7 +3049,7 @@ dsv4-fp4-b300-vllm-agentic:
       - { tp: 8, ep: 8, dp-attn: true, offloading: cpu,  conc-list: [128, 256, 512] }
 
 dsv4-fp4-b300-trt:
-  image: ghcr.io#semianalysisai/trtllm-deepseek-v4:fix-dsv4-swa-scratch-revert-shrink-c914d6d
+  image: ghcr.io#semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-2dd03e6
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: b300
diff --git a/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b200_trt.sh b/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b200_trt.sh
index e4a24dea2..f2b19a8cb 100644
--- a/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b200_trt.sh
+++ b/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b200_trt.sh
@@ -47,6 +47,13 @@ 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 "$MODEL"
 fi
diff --git a/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b300_trt.sh b/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b300_trt.sh
index b0150e10d..3606c677f 100644
--- a/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b300_trt.sh
+++ b/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b300_trt.sh
@@ -59,6 +59,13 @@ sanitize_slurm_mpi_env_for_trtllm
 export NCCL_NVLS_ENABLE="${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"
+
 nvidia-smi
 
 SERVER_LOG="$PWD/server.log"
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index a6258aee1..17d162518 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -3382,7 +3382,7 @@
     - dsv4-fp4-b200-trt
     - dsv4-fp4-b300-trt
   description:
-    - "Update the TensorRT-LLM DeepSeek-V4-Pro image to ghcr.io/semianalysisai/trtllm-deepseek-v4:fix-dsv4-swa-scratch-revert-shrink-c914d6d (TRT-LLM feat/deepseek_v4 @ 084cf2ba + kv_cache_manager_v2 fix: free SWA scratch slots on shrink instead of asserting, which crashed the engine on attention-DP context/generation reverts at high concurrency, e.g. b300 8k1k conc>=512)"
+    - "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."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1636
 
 - config-keys:

From 5adfeb34dde2db0107a848c4b473f56633725375 Mon Sep 17 00:00:00 2001
From: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Date: Wed, 3 Jun 2026 14:04:49 -0700
Subject: [PATCH 7/9] 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>
---
 .github/configs/nvidia-master.yaml                        | 2 +-
 benchmarks/single_node/fixed_seq_len/dsv4_fp4_b200_trt.sh | 7 -------
 perf-changelog.yaml                                       | 3 +--
 3 files changed, 2 insertions(+), 10 deletions(-)

diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index f216c1848..836729dac 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.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
diff --git a/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b200_trt.sh b/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b200_trt.sh
index f2b19a8cb..e4a24dea2 100644
--- a/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b200_trt.sh
+++ b/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
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 0fc1dd2c0..7f19fd14c 100644
--- a/perf-changelog.yaml
+++ b/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:

From ad529fb8be1b526c5940641f8d0f1c4ab37ed1d0 Mon Sep 17 00:00:00 2001
From: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Date: Wed, 3 Jun 2026 14:29:26 -0700
Subject: [PATCH 8/9] Use official TRT-LLM image (1.3.0rc15.post1) for DSv4
 B300 TRT (non-MTP + MTP)

Point dsv4-fp4-b300-trt and dsv4-fp4-b300-trt-mtp at the official
nvcr.io#nvidia/tensorrt-llm/release:1.3.0rc15.post1 (from the custom
feat/deepseek_v4 builds 2dd03e6 / 9aa3715) and drop the
TRTLLM_DSV4_ENABLE_SWA_SCRATCH_REUSE=0 launcher workaround so the official
image runs with native behavior. B200 TRT unchanged (9aa3715).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---
 .github/configs/nvidia-master.yaml                        | 4 ++--
 benchmarks/single_node/fixed_seq_len/dsv4_fp4_b300_trt.sh | 7 -------
 perf-changelog.yaml                                       | 3 ++-
 3 files changed, 4 insertions(+), 10 deletions(-)

diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index 836729dac..fe93ed9a3 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -3049,7 +3049,7 @@ dsv4-fp4-b300-vllm-agentic:
       - { tp: 8, ep: 8, dp-attn: true, offloading: cpu,  conc-list: [128, 256, 512] }
 
 dsv4-fp4-b300-trt:
-  image: ghcr.io#semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-2dd03e6
+  image: nvcr.io#nvidia/tensorrt-llm/release:1.3.0rc15.post1
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: b300
@@ -3072,7 +3072,7 @@ dsv4-fp4-b300-trt:
       - { tp: 8, ep: 8, dp-attn: true, conc-start: 256, conc-end: 1024 }
 
 dsv4-fp4-b300-trt-mtp:
-  image: ghcr.io#semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-9aa3715
+  image: nvcr.io#nvidia/tensorrt-llm/release:1.3.0rc15.post1
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: b300
diff --git a/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b300_trt.sh b/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b300_trt.sh
index 3606c677f..b0150e10d 100644
--- a/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b300_trt.sh
+++ b/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b300_trt.sh
@@ -59,13 +59,6 @@ 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"
-
 nvidia-smi
 
 SERVER_LOG="$PWD/server.log"
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 7f19fd14c..6af5012c4 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -3388,8 +3388,9 @@
 
 - config-keys:
     - dsv4-fp4-b300-trt
+    - dsv4-fp4-b300-trt-mtp
   description:
-    - "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)."
+    - "Point the B300 TensorRT-LLM DeepSeek-V4-Pro configs (non-MTP dsv4-fp4-b300-trt and MTP dsv4-fp4-b300-trt-mtp) at the official NVIDIA release image nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc15.post1, replacing the custom ghcr.io semianalysis feat/deepseek_v4 builds (2dd03e6 and 9aa3715 respectively), to evaluate the official RC for DeepSeek-V4-Pro. Also drops the TRTLLM_DSV4_ENABLE_SWA_SCRATCH_REUSE=0 launcher workaround (specific to the custom build) so the official image runs with its native behavior. B200 TRT is unchanged (stays on feat-deepseek_v4-9aa3715)."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1636
 
 - config-keys:

From c2381b72b1bcfbc275ce6086baded5cc3b1ca17b Mon Sep 17 00:00:00 2001
From: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Date: Thu, 4 Jun 2026 21:44:03 -0700
Subject: [PATCH 9/9] Re-add TRTLLM_DSV4_ENABLE_SWA_SCRATCH_REUSE=0 on
 rc15.post1 to test DPA crash

The previous sweep crashed all dpa=true jobs with CUDA_ERROR_ILLEGAL_ADDRESS
on rc15.post1 without the SWA scratch workaround. Re-add
TRTLLM_DSV4_ENABLE_SWA_SCRATCH_REUSE=0 to both B300 TRT launchers (non-MTP
and MTP) to determine whether the DPA crash is the same SWA-scratch bug or a
separate FMHA kernel issue.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 benchmarks/single_node/fixed_seq_len/dsv4_fp4_b300_trt.sh  | 5 +++++
 .../single_node/fixed_seq_len/dsv4_fp4_b300_trt_mtp.sh     | 5 +++++
 perf-changelog.yaml                                        | 7 +++++++
 3 files changed, 17 insertions(+)

diff --git a/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b300_trt.sh b/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b300_trt.sh
index b0150e10d..35ac41f6b 100644
--- a/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b300_trt.sh
+++ b/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b300_trt.sh
@@ -59,6 +59,11 @@ 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 to test whether the rc15.post1 DPA crash
+# is the same SWA-scratch bug or a separate FMHA kernel issue.
+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"
+
 nvidia-smi
 
 SERVER_LOG="$PWD/server.log"
diff --git a/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b300_trt_mtp.sh b/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b300_trt_mtp.sh
index 507b96e34..f6ac67d6b 100644
--- a/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b300_trt_mtp.sh
+++ b/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b300_trt_mtp.sh
@@ -58,6 +58,11 @@ sanitize_slurm_mpi_env_for_trtllm
 export NCCL_NVLS_ENABLE="${NCCL_NVLS_ENABLE:-0}"
 echo "NCCL_NVLS_ENABLE: $NCCL_NVLS_ENABLE"
 
+# Disable DSv4 SWA scratch reuse to test whether the rc15.post1 DPA crash
+# is the same SWA-scratch bug or a separate FMHA kernel issue.
+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"
+
 nvidia-smi
 
 SERVER_LOG="$PWD/server.log"
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 6af5012c4..41fc63ab1 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -3451,3 +3451,10 @@
     - "Add MiniMax-M2.5 NVFP4 GB300 disaggregated multinode vLLM benchmarks via Dynamo"
     - "Add 1k1k/8k1k minimax recipe set under benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1641
+
+- config-keys:
+    - dsv4-fp4-b300-trt
+    - dsv4-fp4-b300-trt-mtp
+  description:
+    - "Switch B300 DSv4 TRT (non-MTP + MTP) to official rc15.post1 image with TRTLLM_DSV4_ENABLE_SWA_SCRATCH_REUSE=0 to test whether DPA crash is same SWA-scratch bug"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1636