fix(amd): preserve DSV4 MoRI dispatch pin

Keep the proven 128-token per-rank decode dispatch setting for both DEP8 fixed-sequence recipes while the shared launcher derives only admission and the inter-kernel threshold.

Author: cquil11

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/mi355x/fixed-seq/disagg-1p1d-dep8-mtp.yaml

@@ -105,7 +105,7 @@ backend:
     SGLANG_DISAGGREGATION_FORCE_QUERY_PREFILL_DP_RANK: "1"
   decode_environment:
     <<: *common_environment
-    SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK: "1024"
+    SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK: "128"
   sglang_config:
     prefill: &common_server
       served-model-name: "deepseek-ai/DeepSeek-V4-Pro"

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/mi355x/fixed-seq/disagg-1p1d-dep8-stp.yaml

@@ -105,7 +105,7 @@ backend:
     SGLANG_DISAGGREGATION_FORCE_QUERY_PREFILL_DP_RANK: "1"
   decode_environment:
     <<: *common_environment
-    SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK: "1024"
+    SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK: "128"
   sglang_config:
     prefill: &common_server
       served-model-name: "deepseek-ai/DeepSeek-V4-Pro"