Keep SGLANG_RADIX_DISABLE_REUSE in dsv4 8k1k recipes

Restore SGLANG_RADIX_DISABLE_REUSE: "1" in both the prefill and decode
environment blocks of all 6 recipes (right after PYTHONUNBUFFERED), reverting
that part of the #1559-style cleanup; update the changelog accordingly.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: Fridge003

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-low-latency-1p1d-tp4-tp4-mtp.yaml

@@ -31,6 +31,7 @@ backend:
 
   prefill_environment:
     PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_DISABLE_REUSE: "1"
     SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
     SGLANG_DEFAULT_THINKING: "1"
     SGLANG_DSV4_REASONING_EFFORT: "max"
@@ -45,6 +46,7 @@ backend:
 
   decode_environment:
     PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_DISABLE_REUSE: "1"
     SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
     SGLANG_DEFAULT_THINKING: "1"
     SGLANG_DSV4_REASONING_EFFORT: "max"

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-low-latency-1p6d-dep4-tp4-mtp.yaml

@@ -31,6 +31,7 @@ backend:
 
   prefill_environment:
     PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_DISABLE_REUSE: "1"
     SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
     SGLANG_DEFAULT_THINKING: "1"
     SGLANG_DSV4_REASONING_EFFORT: "max"
@@ -52,6 +53,7 @@ backend:
 
   decode_environment:
     PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_DISABLE_REUSE: "1"
     SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
     SGLANG_DEFAULT_THINKING: "1"
     SGLANG_DSV4_REASONING_EFFORT: "max"

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-mid-curve-1p1d-dep4-dep16-mtp.yaml

@@ -33,6 +33,7 @@ backend:
 
   prefill_environment:
     PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_DISABLE_REUSE: "1"
     SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
     SGLANG_DEFAULT_THINKING: "1"
     SGLANG_DSV4_REASONING_EFFORT: "max"
@@ -54,6 +55,7 @@ backend:
 
   decode_environment:
     PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_DISABLE_REUSE: "1"
     SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
     SGLANG_DEFAULT_THINKING: "1"
     SGLANG_DSV4_REASONING_EFFORT: "max"

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-mid-curve-1p1d-dep4-dep8-mtp.yaml

@@ -33,6 +33,7 @@ backend:
 
   prefill_environment:
     PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_DISABLE_REUSE: "1"
     SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
     SGLANG_DEFAULT_THINKING: "1"
     SGLANG_DSV4_REASONING_EFFORT: "max"
@@ -54,6 +55,7 @@ backend:
 
   decode_environment:
     PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_DISABLE_REUSE: "1"
     SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
     SGLANG_DEFAULT_THINKING: "1"
     SGLANG_DSV4_REASONING_EFFORT: "max"

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-mid-curve-2p1d-dep4-dep8-mtp.yaml

@@ -33,6 +33,7 @@ backend:
 
   prefill_environment:
     PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_DISABLE_REUSE: "1"
     SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
     SGLANG_DEFAULT_THINKING: "1"
     SGLANG_DSV4_REASONING_EFFORT: "max"
@@ -54,6 +55,7 @@ backend:
 
   decode_environment:
     PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_DISABLE_REUSE: "1"
     SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
     SGLANG_DEFAULT_THINKING: "1"
     SGLANG_DSV4_REASONING_EFFORT: "max"

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-mid-curve-4p1d-dep4-dep8-mtp.yaml

@@ -33,6 +33,7 @@ backend:
 
   prefill_environment:
     PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_DISABLE_REUSE: "1"
     SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
     SGLANG_DEFAULT_THINKING: "1"
     SGLANG_DSV4_REASONING_EFFORT: "max"
@@ -52,6 +53,7 @@ backend:
 
   decode_environment:
     PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_DISABLE_REUSE: "1"
     SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
     SGLANG_DEFAULT_THINKING: "1"
     SGLANG_DSV4_REASONING_EFFORT: "max"

perf-changelog.yaml

@@ -3116,7 +3116,7 @@
     - "Enable W4A4 (MXFP4) megamoe by setting SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS=1 and SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND=1 in the megamoe environment blocks (not the low-latency 1p1d-tp4-tp4 or the 4p1d recipe)"
     - "Update SGLang image from nightly-dev-cu13-20260510-2473659e to nightly-dev-20260527-14f81a67"
     - "Switch moe-a2a-backend from deepep to megamoe and drop the deepep-config override"
-    - "Clean up obsolete environs in the 8k1k disagg recipes: drop SGLANG_OPT_USE_JIT_NORM / SGLANG_OPT_USE_JIT_INDEXER_METADATA / SGLANG_OPT_USE_TOPK_V2 (now default-on); drop the auto-set MegaMoE companions (SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE, SGLANG_OPT_FIX_HASH_MEGA_MOE, SGLANG_OPT_FIX_MEGA_MOE_MEMORY, SGLANG_OPT_FIX_NEXTN_MEGA_MOE, SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK); drop SGLANG_RADIX_DISABLE_REUSE / SGLANG_OPT_USE_FAST_MASK_EP which no longer exist in sglang environ.py"
+    - "Clean up obsolete environs in the 8k1k disagg recipes: drop SGLANG_OPT_USE_JIT_NORM / SGLANG_OPT_USE_JIT_INDEXER_METADATA / SGLANG_OPT_USE_TOPK_V2 (now default-on); drop the auto-set MegaMoE companions (SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE, SGLANG_OPT_FIX_HASH_MEGA_MOE, SGLANG_OPT_FIX_MEGA_MOE_MEMORY, SGLANG_OPT_FIX_NEXTN_MEGA_MOE, SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK); drop SGLANG_OPT_USE_FAST_MASK_EP which no longer exists in sglang environ.py (SGLANG_RADIX_DISABLE_REUSE is kept)"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1378
 
 - config-keys: