[testing][nv] h100: mtp configs sglang

.github/configs/nvidia-master.yaml

@@ -2976,29 +2976,59 @@ dsr1-fp8-h100-dynamo-sglang:
   - isl: 1024
     osl: 1024
     search-space:
-    # Max throughput TEP (1 prefill, 2 decode)
-    - conc-list: [1, 2, 4, 8, 16, 32, 64, 128]
+    # # STP: Max throughput TEP (1 prefill, 2 decode)
+    # - conc-list: [1, 2, 4, 8, 16, 32, 64, 128]
+    #   prefill:
+    #     num-worker: 1
+    #     tp: 16
+    #     ep: 1
+    #     dp-attn: false
+    #     additional-settings:
+    #     - "CONFIG_FILE=recipes/h100/1k1k/stp/h100-fp8-1p2d-max-tp.yaml"
+    #   decode:
+    #     num-worker: 2
+    #     tp: 16
+    #     ep: 1
+    #     dp-attn: false
+    # # STP: Max throughput DEP (1 prefill, 1 decode, dp-attention)
+    # - conc-list: [1, 2, 4, 8, 16, 32, 64]
+    #   prefill:
+    #     num-worker: 1
+    #     tp: 16
+    #     ep: 1
+    #     dp-attn: false
+    #     additional-settings:
+    #     - "CONFIG_FILE=recipes/h100/1k1k/stp/h100-fp8-1p1d-max-dep.yaml"
+    #   decode:
+    #     num-worker: 1
+    #     tp: 16
+    #     ep: 16
+    #     dp-attn: true
+    # MTP: Max throughput TEP (1 prefill, 2 decode)
+    - spec-decoding: "mtp"
+      conc-list: [1, 2, 4, 8, 16, 32, 64, 128]
       prefill:
         num-worker: 1
         tp: 16
         ep: 1
         dp-attn: false
         additional-settings:
-        - "CONFIG_FILE=recipes/h100/1k1k/stp/h100-fp8-1p2d-max-tp.yaml"
+        - "CONFIG_FILE=recipes/h100/1k1k/mtp/h100-fp8-1p2d-max-tp-mtp.yaml"
       decode:
         num-worker: 2
         tp: 16
         ep: 1
         dp-attn: false
-    # Max throughput DEP (1 prefill, 1 decode, dp-attention)
-    - conc-list: [1, 2, 4, 8, 16, 32, 64]
+    # MTP: Max throughput DEP (1 prefill, 1 decode, dp-attention)
+    - spec-decoding: "mtp"
+      conc-list: [1, 2, 4, 8, 16, 32, 64]
       prefill:
         num-worker: 1
         tp: 16
         ep: 1
         dp-attn: false
         additional-settings:
-        - "CONFIG_FILE=recipes/h100/1k1k/stp/h100-fp8-1p1d-max-dep.yaml"
+        - "CONFIG_FILE=recipes/h100/1k1k/mtp/h100-fp8-1p1d-max-dep-mtp.yaml"
       decode:
         num-worker: 1
         tp: 16
@@ -3007,29 +3037,59 @@ dsr1-fp8-h100-dynamo-sglang:
   - isl: 1024
     osl: 8192
     search-space:
-    # Max throughput TEP (1 prefill, 2 decode)
-    - conc-list: [1, 2, 4, 8, 16, 32]
+    # # STP: Max throughput TEP (1 prefill, 2 decode)
+    # - conc-list: [1, 2, 4, 8, 16, 32]
+    #   prefill:
+    #     num-worker: 1
+    #     tp: 16
+    #     ep: 1
+    #     dp-attn: false
+    #     additional-settings:
+    #     - "CONFIG_FILE=recipes/h100/1k8k/stp/h100-fp8-1p2d-max-tp.yaml"
+    #   decode:
+    #     num-worker: 2
+    #     tp: 16
+    #     ep: 1
+    #     dp-attn: false
+    # # STP: Max throughput DEP (1 prefill, 1 decode, dp-attention)
+    # - conc-list: [1, 2, 4, 8]
+    #   prefill:
+    #     num-worker: 1
+    #     tp: 16
+    #     ep: 1
+    #     dp-attn: false
+    #     additional-settings:
+    #     - "CONFIG_FILE=recipes/h100/1k8k/stp/h100-fp8-1p1d-max-dep.yaml"
+    #   decode:
+    #     num-worker: 1
+    #     tp: 16
+    #     ep: 16
+    #     dp-attn: true
+    # MTP: Max throughput TEP (1 prefill, 2 decode)
+    - spec-decoding: "mtp"
+      conc-list: [1, 2, 4, 8, 16, 32, 64]
       prefill:
         num-worker: 1
         tp: 16
         ep: 1
         dp-attn: false
         additional-settings:
-        - "CONFIG_FILE=recipes/h100/1k8k/stp/h100-fp8-1p2d-max-tp.yaml"
+        - "CONFIG_FILE=recipes/h100/1k8k/mtp/h100-fp8-1p2d-max-tp-mtp.yaml"
       decode:
         num-worker: 2
         tp: 16
         ep: 1
         dp-attn: false
-    # Max throughput DEP (1 prefill, 1 decode, dp-attention)
-    - conc-list: [1, 2, 4, 8]
+    # MTP: Max throughput DEP (1 prefill, 1 decode, dp-attention)
+    - spec-decoding: "mtp"
+      conc-list: [1, 2, 4, 8, 16, 32, 64]
       prefill:
         num-worker: 1
         tp: 16
         ep: 1
         dp-attn: false
         additional-settings:
-        - "CONFIG_FILE=recipes/h100/1k8k/stp/h100-fp8-1p1d-max-dep.yaml"
+        - "CONFIG_FILE=recipes/h100/1k8k/mtp/h100-fp8-1p1d-max-dep-mtp.yaml"
       decode:
         num-worker: 1
         tp: 16
@@ -3038,29 +3098,59 @@ dsr1-fp8-h100-dynamo-sglang:
   - isl: 8192
     osl: 1024
     search-space:
-    # Max throughput TEP (1 prefill, 1 decode)
-    - conc-list: [1, 2, 4, 8, 16, 32, 64, 128]
+    # # STP: Max throughput TEP (1 prefill, 1 decode)
+    # - conc-list: [1, 2, 4, 8, 16, 32, 64, 128]
+    #   prefill:
+    #     num-worker: 1
+    #     tp: 16
+    #     ep: 1
+    #     dp-attn: false
+    #     additional-settings:
+    #     - "CONFIG_FILE=recipes/h100/8k1k/stp/h100-fp8-1p1d-max-tp.yaml"
+    #   decode:
+    #     num-worker: 1
+    #     tp: 16
+    #     ep: 1
+    #     dp-attn: false
+    # # STP: Max throughput DEP (1 prefill, 1 decode, dp-attention)
+    # - conc-list: [1, 2, 4, 8, 16, 32, 64]
+    #   prefill:
+    #     num-worker: 1
+    #     tp: 16
+    #     ep: 1
+    #     dp-attn: false
+    #     additional-settings:
+    #     - "CONFIG_FILE=recipes/h100/8k1k/stp/h100-fp8-1p1d-max-dep.yaml"
+    #   decode:
+    #     num-worker: 1
+    #     tp: 16
+    #     ep: 16
+    #     dp-attn: true
+    # MTP: Max throughput TEP (1 prefill, 1 decode)
+    - spec-decoding: "mtp"
+      conc-list: [1, 2, 4, 8, 16, 32, 64, 128]
       prefill:
         num-worker: 1
         tp: 16
         ep: 1
         dp-attn: false
         additional-settings:
-        - "CONFIG_FILE=recipes/h100/8k1k/stp/h100-fp8-1p1d-max-tp.yaml"
+        - "CONFIG_FILE=recipes/h100/8k1k/mtp/h100-fp8-1p1d-max-tp-mtp.yaml"
       decode:
         num-worker: 1
         tp: 16
         ep: 1
         dp-attn: false
-    # Max throughput DEP (1 prefill, 1 decode, dp-attention)
-    - conc-list: [1, 2, 4, 8, 16, 32, 64]
+    # MTP: Max throughput DEP (1 prefill, 1 decode, dp-attention)
+    - spec-decoding: "mtp"
+      conc-list: [1, 2, 4, 8, 16, 32, 64]
       prefill:
         num-worker: 1
         tp: 16
         ep: 1
         dp-attn: false
         additional-settings:
-        - "CONFIG_FILE=recipes/h100/8k1k/stp/h100-fp8-1p1d-max-dep.yaml"
+        - "CONFIG_FILE=recipes/h100/8k1k/mtp/h100-fp8-1p1d-max-dep-mtp.yaml"
       decode:
         num-worker: 1
         tp: 16

perf-changelog.yaml

@@ -494,6 +494,24 @@
     - "9 recipes: 4x 1k1k + 5x 8k1k, low-latency and max-throughput with EAGLE speculative decoding"
   pr-link: https://github.com/InferenceMAX/InferenceMAX/pull/667
 
+- config-keys:
+    - dsr1-fp8-h100-dynamo-sglang
+  description:
+    - "Add DeepSeek-R1 FP8 H100 Dynamo SGLang STP disaggregated multinode configurations"
+    - "Image: lmsysorg/sglang:v0.5.8-cu130"
+    - "1k1k, 1k8k, 8k1k sequence lengths"
+    - "Two modes per seq-len: Max throughput TEP (1P2D) and Max throughput DEP (1P1D with dp-attention)"
+  pr-link: https://github.com/InferenceMAX/InferenceMAX/pull/643
+
+- config-keys:
+    - dsr1-fp8-h100-dynamo-sglang
+  description:
+    - "Add DeepSeek-R1 FP8 H100 Dynamo SGLang MTP disaggregated multinode configurations"
+    - "Image: lmsysorg/sglang:v0.5.8-cu130"
+    - "1k1k, 1k8k, 8k1k sequence lengths with MTP speculative decoding"
+    - "Two modes per seq-len: Max throughput TEP (1P2D) and Max throughput DEP (1P1D with dp-attention)"
+  pr-link: https://github.com/InferenceMAX/InferenceMAX/pull/644
+
 - config-keys:
     - dsr1-fp8-mi355x-atom-mtp
     - dsr1-fp4-mi355x-atom-mtp
@@ -527,14 +545,14 @@
     - "Add conc 4/8/16"
     - "Use Pure TP with MTP=2 for 1k1k conc smaller than 128 and reduce MTP to 1 for DEP configs"
   pr-link: https://github.com/InferenceMAX/InferenceMAX/pull/674
-  
+
 - config-keys:
     - dsr1-fp4-b200-dynamo-sglang-mtp
   description:
     - "Add B200 configs for DSR1 FP4 SGLANG MTP mode for 1k1k and 8k1k"
     - "Image: lmsysorg/sglang:v0.5.8.post1-cu130"
   pr-link: https://github.com/InferenceMAX/InferenceMAX/pull/683
-  
+
 - config-keys:
     - dsr1-fp4-gb300-dynamo-sglang
   description:
@@ -564,14 +582,14 @@
   description:
     - "Add more sweep points for DSR1 FP8 both MTP and non-MTP 1k1k, 8k1k"
   pr-link: https://github.com/InferenceMAX/InferenceMAX/pull/689
-  
+
 - config-keys:
     - dsr1-fp8-b200-dynamo-trt
   description:
     - "Update max_num_tokens and max_batch_size for min-latency decode workers"
     - "See srt-slurm recipe changes: https://github.com/ishandhanani/srt-slurm/pull/173"
   pr-link: https://github.com/InferenceMAX/InferenceMAX/pull/686
-  
+
 - config-keys:
     - dsr1-fp8-mi325x-sglang
   description:

runners/launch_h100-dgxc-slurm.sh

@@ -81,7 +81,7 @@ cat > srtslurm.yaml <<EOF
 # Default SLURM settings
 default_account: "${SLURM_ACCOUNT}"
 default_partition: "${SLURM_PARTITION}"
-default_time_limit: "4:00:00"
+default_time_limit: "6:00:00"
 # Resource defaults
 gpus_per_node: 8
 network_interface: ""