Add config file for fused MoE for Nemotron (TP4, B200) (#34411)

Signed-off-by: Daniel Serebrenik <daserebrenik@nvidia.com>

Add config file for fused MoE for Nemotron (TP4, B200) (#34411)
Signed-off-by: Daniel Serebrenik <daserebrenik@nvidia.com>
dea63512 · danisereb · GitHub · 8a798be9 · dea63512
Unverified Commit dea63512 authored Feb 12, 2026 by danisereb Committed by GitHub Feb 12, 2026
Hide whitespace changes
Inline Side-by-side

Showing with 59 additions and 0 deletions

vllm/model_executor/layers/fused_moe/configs/E=512,N=672,device_name=NVIDIA_B200.json ...used_moe/configs/E=512,N=672,device_name=NVIDIA_B200.json +59 -0

No files found.
--- a/vllm/model_executor/layers/fused_moe/configs/E=512,N=672,device_name=NVIDIA_B200.json
+++ b/vllm/model_executor/layers/fused_moe/configs/E=512,N=672,device_name=NVIDIA_B200.json
+{
+    "triton_version": "3.6.0",
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "256": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "512": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2
+    }
+}