fix: switch gb200 sglang mtp recipes to nixl transfer backend

mooncake's MNNVL handshake fails on every prefill->decode transfer on the
GB200 NV cluster, producing 0 output tokens with all requests aborted
(KVTransferError(...): Aborted by AbortReq, ~100K errors/job in run
25785003012). Same dynamo hash + same SGLang container with dep8-dep8
works on GB300, isolating the break to GB200's MNNVL fabric. Switching to
nixl uses dynamo's native transport, which has GB200 NVL4 support.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Author: Oseltamivir

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-low-latency-1p1d-tp8-tp8-mtp.yaml

@@ -83,7 +83,7 @@ backend:
       tool-call-parser: deepseekv4  # gates dsv4 chat-encoding spec.
 
       disaggregation-mode: "prefill"
-      disaggregation-transfer-backend: mooncake
+      disaggregation-transfer-backend: nixl
 
       tensor-parallel-size: 8
       data-parallel-size: 1
@@ -104,7 +104,7 @@ backend:
       tool-call-parser: deepseekv4  # gates dsv4 chat-encoding spec.
 
       disaggregation-mode: "decode"
-      disaggregation-transfer-backend: mooncake
+      disaggregation-transfer-backend: nixl
 
       tensor-parallel-size: 8
       data-parallel-size: 1

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-low-latency-1p6d-dep8-tp8-mtp.yaml

@@ -94,7 +94,7 @@ backend:
       tool-call-parser: deepseekv4  # gates dsv4 chat-encoding spec.
 
       disaggregation-mode: "prefill"
-      disaggregation-transfer-backend: mooncake
+      disaggregation-transfer-backend: nixl
 
       tensor-parallel-size: 8
       data-parallel-size: 8
@@ -118,7 +118,7 @@ backend:
       tool-call-parser: deepseekv4  # gates dsv4 chat-encoding spec.
 
       disaggregation-mode: "decode"
-      disaggregation-transfer-backend: mooncake
+      disaggregation-transfer-backend: nixl
 
       tensor-parallel-size: 8
       data-parallel-size: 1

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-1p1d-dep8-dep16-mtp.yaml

@@ -103,7 +103,7 @@ backend:
       tool-call-parser: deepseekv4  # gates dsv4 chat-encoding spec.
 
       disaggregation-mode: "prefill"
-      disaggregation-transfer-backend: mooncake
+      disaggregation-transfer-backend: nixl
 
       tensor-parallel-size: 8
       data-parallel-size: 8
@@ -128,7 +128,7 @@ backend:
       tool-call-parser: deepseekv4  # gates dsv4 chat-encoding spec.
 
       disaggregation-mode: "decode"
-      disaggregation-transfer-backend: mooncake
+      disaggregation-transfer-backend: nixl
 
       tensor-parallel-size: 16
       data-parallel-size: 16

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-1p1d-dep8-dep8-mtp.yaml

@@ -103,7 +103,7 @@ backend:
       tool-call-parser: deepseekv4  # gates dsv4 chat-encoding spec.
 
       disaggregation-mode: "prefill"
-      disaggregation-transfer-backend: mooncake
+      disaggregation-transfer-backend: nixl
 
       tensor-parallel-size: 8
       data-parallel-size: 8
@@ -128,7 +128,7 @@ backend:
       tool-call-parser: deepseekv4  # gates dsv4 chat-encoding spec.
 
       disaggregation-mode: "decode"
-      disaggregation-transfer-backend: mooncake
+      disaggregation-transfer-backend: nixl
 
       tensor-parallel-size: 8
       data-parallel-size: 8

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-4p1d-dep8-dep8-mtp-c8192.yaml

@@ -103,7 +103,7 @@ backend:
       tool-call-parser: deepseekv4  # gates dsv4 chat-encoding spec.
 
       disaggregation-mode: "prefill"
-      disaggregation-transfer-backend: mooncake
+      disaggregation-transfer-backend: nixl
 
       tensor-parallel-size: 8
       data-parallel-size: 8
@@ -128,7 +128,7 @@ backend:
       tool-call-parser: deepseekv4  # gates dsv4 chat-encoding spec.
 
       disaggregation-mode: "decode"
-      disaggregation-transfer-backend: mooncake
+      disaggregation-transfer-backend: nixl
 
       tensor-parallel-size: 8
       data-parallel-size: 8

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-5p1d-dep8-dep8-mtp-c12288.yaml

@@ -103,7 +103,7 @@ backend:
       tool-call-parser: deepseekv4  # gates dsv4 chat-encoding spec.
 
       disaggregation-mode: "prefill"
-      disaggregation-transfer-backend: mooncake
+      disaggregation-transfer-backend: nixl
 
       tensor-parallel-size: 8
       data-parallel-size: 8
@@ -128,7 +128,7 @@ backend:
       tool-call-parser: deepseekv4  # gates dsv4 chat-encoding spec.
 
       disaggregation-mode: "decode"
-      disaggregation-transfer-backend: mooncake
+      disaggregation-transfer-backend: nixl
 
       tensor-parallel-size: 8
       data-parallel-size: 8

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-6p1d-dep8-dep8-mtp-c16384.yaml

@@ -103,7 +103,7 @@ backend:
       tool-call-parser: deepseekv4  # gates dsv4 chat-encoding spec.
 
       disaggregation-mode: "prefill"
-      disaggregation-transfer-backend: mooncake
+      disaggregation-transfer-backend: nixl
 
       tensor-parallel-size: 8
       data-parallel-size: 8
@@ -128,7 +128,7 @@ backend:
       tool-call-parser: deepseekv4  # gates dsv4 chat-encoding spec.
 
       disaggregation-mode: "decode"
-      disaggregation-transfer-backend: mooncake
+      disaggregation-transfer-backend: nixl
 
       tensor-parallel-size: 8
       data-parallel-size: 8