[NV] update B300 disagg recipes (SemiAnalysisAI#1891)

* [NV] Add MiniMax M3 B300 Dynamo vLLM recipes

* chore: update MiniMax M3 B300 container

* chore: update changelog PR link

* Update perf-changelog.yaml

* Update perf-changelog.yaml

* fix(vllm): patch MiniMax M3 MSA contiguity

* fix(recipes): align MiniMax M3 parallel settings

* fix(vllm): backport MiniMax M3 eval fixes

* ci(sweep): enable full MiniMax M3 validation

* perf(vllm): right-size MiniMax M3 low concurrency

* perf(vllm): colocate MiniMax M3 TP4 workers

* fix(runner): exclude faulty B300 RDMA node

* fix(runner): verify B300 node exclusion

* fix(runner): check generated B300 sbatch script

* ci(sweep): validate B300 node exclusion

* refactor(vllm): trim MiniMax M3 runtime patches

* Update MiniMax M3 B300 Dynamo vLLM recipes

* fix

* update to flashinfer

* prune non-pareto

* fix(vllm): remove pruned MiniMax M3 B300 recipes

---------

Co-authored-by: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Co-authored-by: Biswa Panda <biswa.panda@gmail.com>
Co-authored-by: Ankur Singh <ankusingh@nvidia.com>
Co-authored-by: Rohit Nagraj <rohitnagraj.99@gmail.com>

Author: 5 people

.github/configs/nvidia-master.yaml

@@ -11850,19 +11850,6 @@ minimaxm3-fp8-b300-dynamo-vllm:
     - isl: 8192
       osl: 1024
       search-space:
-      - conc-list: [128]
-        prefill:
-          num-worker: 1
-          tp: 2
-          ep: 2
-          dp-attn: true
-          additional-settings:
-          - "CONFIG_FILE=recipes/vllm/minimax-m3/b300-fp8/8k1k/1p2d-dep2-dep8-8k1k.yaml"
-        decode:
-          num-worker: 2
-          tp: 8
-          ep: 8
-          dp-attn: true
       - conc-list: [256, 512]
         prefill:
           num-worker: 2
@@ -11889,31 +11876,31 @@ minimaxm3-fp8-b300-dynamo-vllm:
           tp: 8
           ep: 8
           dp-attn: false
-      - conc-list: [512]
+      - conc-list: [4096]
         prefill:
-          num-worker: 3
+          num-worker: 4
           tp: 2
           ep: 2
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/vllm/minimax-m3/b300-fp8/8k1k/3p2d-dep2-dep8-8k1k.yaml"
+          - "CONFIG_FILE=recipes/vllm/minimax-m3/b300-fp8/8k1k/4p2d-dep2-dep8-8k1k.yaml"
         decode:
           num-worker: 2
           tp: 8
           ep: 8
           dp-attn: true
-      - conc-list: [32]
+      - conc-list: [1, 4, 8, 16]
         prefill:
-          num-worker: 3
+          num-worker: 1
           tp: 2
           ep: 2
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/vllm/minimax-m3/b300-fp8/8k1k/3p2d-dep2-tep8-8k1k.yaml"
+          - "CONFIG_FILE=recipes/vllm/minimax-m3/b300-fp8/8k1k/1p1d-dep2-tp4-marlin-8k1k.yaml"
         decode:
-          num-worker: 2
-          tp: 8
-          ep: 8
+          num-worker: 1
+          tp: 4
+          ep: 1
           dp-attn: false
       - conc-list: [4096]
         prefill:
@@ -11922,50 +11909,50 @@ minimaxm3-fp8-b300-dynamo-vllm:
           ep: 2
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/vllm/minimax-m3/b300-fp8/8k1k/4p2d-dep2-dep8-8k1k.yaml"
+          - "CONFIG_FILE=recipes/vllm/minimax-m3/b300-fp8/8k1k/4p2d-dep2-tep4-8k1k.yaml"
         decode:
           num-worker: 2
-          tp: 8
-          ep: 8
-          dp-attn: true
-      - conc-list: [4096]
+          tp: 4
+          ep: 4
+          dp-attn: false
+      - conc-list: [16, 32, 64, 128]
         prefill:
-          num-worker: 4
+          num-worker: 1
           tp: 2
           ep: 2
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/vllm/minimax-m3/b300-fp8/8k1k/4p3d-dep2-dep4-8k1k.yaml"
+          - "CONFIG_FILE=recipes/vllm/minimax-m3/b300-fp8/8k1k/1p4d-dep2-tep4-8k1k.yaml"
         decode:
-          num-worker: 3
+          num-worker: 4
           tp: 4
           ep: 4
-          dp-attn: true
-      - conc-list: [4, 64]
+          dp-attn: false
+      - conc-list: [16]
         prefill:
-          num-worker: 5
+          num-worker: 1
           tp: 2
           ep: 2
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/vllm/minimax-m3/b300-fp8/8k1k/5p2d-dep2-tep8-8k1k.yaml"
+          - "CONFIG_FILE=recipes/vllm/minimax-m3/b300-fp8/8k1k/1p2d-dep2-tep4-8k1k.yaml"
         decode:
           num-worker: 2
-          tp: 8
-          ep: 8
+          tp: 4
+          ep: 4
           dp-attn: false
-      - conc-list: [1, 4, 8, 16]
+      - conc-list: [4]
         prefill:
           num-worker: 1
           tp: 2
           ep: 2
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/vllm/minimax-m3/b300-fp8/8k1k/1p1d-dep2-tp4-marlin-8k1k.yaml"
+          - "CONFIG_FILE=recipes/vllm/minimax-m3/b300-fp8/8k1k/1p4d-dep2-tep8-8k1k.yaml"
         decode:
-          num-worker: 1
-          tp: 4
-          ep: 1
+          num-worker: 4
+          tp: 8
+          ep: 8
           dp-attn: false
 
 # MiniMax-M3 GB300 disagg sweep — adapted from NV B300 PR #1863.

benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/1p1d-dep2-tp4-marlin-8k1k.yaml

@@ -45,6 +45,8 @@ backend:
       trust-remote-code: true
       no-enable-prefix-caching: true
       kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      kv-cache-dtype: fp8
       block-size: 128
       gpu-memory-utilization: 0.90
       max-model-len: 9472
@@ -60,6 +62,8 @@ backend:
       trust-remote-code: true
       no-enable-prefix-caching: true
       kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      kv-cache-dtype: fp8
       block-size: 128
       gpu-memory-utilization: 0.90
       max-model-len: 9472

benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/1p2d-dep2-tep4-8k1k.yaml

@@ -0,0 +1,83 @@
+name: "minimax-m3-vllm-disagg-b300-1p2d-fp8-dep2-tep4-8k1k"
+
+model:
+  path: "MiniMaxAI/MiniMax-M3-MXFP8"
+  container: "vllm/vllm-openai:minimax-m3-0618-x86_64-cu130"
+  precision: "fp8"
+
+resources:
+  gpu_type: "b300"
+  gpus_per_node: 8
+  prefill_nodes: 1
+  decode_nodes: 1
+  prefill_workers: 1
+  decode_workers: 2
+  gpus_per_prefill: 2
+  gpus_per_decode: 4
+
+dynamo:
+  install: true
+  version: 1.3.0.dev20260614
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+
+  prefill_environment:
+    VLLM_FLOAT32_MATMUL_PRECISION: high
+    UCX_TLS: "cuda_copy,rc"
+
+  decode_environment:
+    VLLM_FLOAT32_MATMUL_PRECISION: high
+    UCX_TLS: "cuda_copy,rc"
+
+  vllm_config:
+    prefill:
+      tensor-parallel-size: 1
+      data-parallel-size: 2
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      kv-cache-dtype: fp8
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 9472
+      language-model-only: true
+      stream-interval: 32
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 16384
+
+    decode:
+      tensor-parallel-size: 4
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      kv-cache-dtype: fp8
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 9472
+      language-model-only: true
+      stream-interval: 32
+      max-num-seqs: 512
+      max-num-batched-tokens: 16384
+      max-cudagraph-capture-size: 4096
+
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "16"
+  req_rate: "inf"

benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/1p4d-dep2-tep4-8k1k.yaml

@@ -0,0 +1,83 @@
+name: "minimax-m3-vllm-disagg-b300-1p4d-fp8-dep2-tep4-8k1k"
+
+model:
+  path: "MiniMaxAI/MiniMax-M3-MXFP8"
+  container: "vllm/vllm-openai:minimax-m3-0618-x86_64-cu130"
+  precision: "fp8"
+
+resources:
+  gpu_type: "b300"
+  gpus_per_node: 8
+  prefill_nodes: 1
+  decode_nodes: 2
+  prefill_workers: 1
+  decode_workers: 4
+  gpus_per_prefill: 2
+  gpus_per_decode: 4
+
+dynamo:
+  install: true
+  version: 1.3.0.dev20260614
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+
+  prefill_environment:
+    VLLM_FLOAT32_MATMUL_PRECISION: high
+    UCX_TLS: "cuda_copy,rc"
+
+  decode_environment:
+    VLLM_FLOAT32_MATMUL_PRECISION: high
+    UCX_TLS: "cuda_copy,rc"
+
+  vllm_config:
+    prefill:
+      tensor-parallel-size: 1
+      data-parallel-size: 2
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      kv-cache-dtype: fp8
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 9472
+      language-model-only: true
+      stream-interval: 32
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 16384
+
+    decode:
+      tensor-parallel-size: 4
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      kv-cache-dtype: fp8
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 9472
+      language-model-only: true
+      stream-interval: 32
+      max-num-seqs: 512
+      max-num-batched-tokens: 16384
+      max-cudagraph-capture-size: 4096
+
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "16x32x64x128"
+  req_rate: "inf"