[Rubin] Enable NVLink all-reduce paths on SM107 (#49647)

Signed-off-by: Zachary Aristei <zaristei@nvidia.com>
Co-authored-by: Sahithi Chigurupati <chigurupati.sahithi@gmail.com>
Co-authored-by: hlu1 <14827759+hlu1@users.noreply.github.com>

vllm/compilation/passes/fusion/allreduce_rms_fusion.py

@@ -119,6 +119,11 @@ FI_ALLREDUCE_FUSION_MAX_SIZE_MB: dict[int, dict[int, float]] = {
         8: 2,  # 2MB
         16: 64,  # 64MB (mnnvl multi-node)
     },
+    107: {
+        2: 64,  # 64MB
+        4: 64,  # 64MB
+        8: 2,  # 2MB
+    },
 }
 
 # Max size of the input tensor per world size per device capability
@@ -140,6 +145,11 @@ _FI_ALLREDUCE_ONE_SHOT_MAX_SIZES_MB: dict[int, dict[int, float]] = {
         4: 4,  # 4MB
         8: 2,  # 2MB
     },
+    107: {
+        2: 32,  # 32MB
+        4: 4,  # 4MB
+        8: 2,  # 2MB
+    },
 }
 
 MiB = 1024 * 1024

vllm/distributed/device_communicators/all_reduce_utils.py

@@ -47,6 +47,12 @@ CUSTOM_ALL_REDUCE_MAX_SIZES = {
         6: 8 * MiB,  # 8 MB
         8: 4 * MiB,  # 4 MB
     },
+    "10.7": {  # sm_107 (Rubin): reuse 10.3 all-reduce thresholds
+        2: 4 * MiB,  # 4 MB
+        4: 4 * MiB,  # 4 MB
+        6: 8 * MiB,  # 8 MB
+        8: 4 * MiB,  # 4 MB
+    },
 }
 
 SYMM_MEM_ALL_REDUCE_MAX_SIZES = {
@@ -68,6 +74,12 @@ SYMM_MEM_ALL_REDUCE_MAX_SIZES = {
         6: 32 * MiB,  # 32 MB
         8: 64 * MiB,  # 64 MB
     },
+    "10.7": {  # sm_107 (Rubin): reuse 10.3 all-reduce thresholds
+        2: 4 * MiB,  # 4 MB
+        4: 32 * MiB,  # 32 MB
+        6: 32 * MiB,  # 32 MB
+        8: 64 * MiB,  # 64 MB
+    },
 }
 
 # NCCL symmetric memory allreduce configuration based on H100 and GB200 benchmarks.

vllm/distributed/device_communicators/symm_mem.py

@@ -27,6 +27,7 @@ class SymmMemCommunicator:
         "9.0": [4, 6, 8],
         "10.0": [6, 8],
         "10.3": [6, 8],
+        "10.7": [6, 8],  # sm_107 (Rubin): reuse 10.3 thresholds
     }
 
     def __init__(