Fix int4 MoE tuner config filename (#35260)

Signed-off-by: maithilijoshi20 <maithilij2003@gmail.com>

benchmark/kernels/fused_moe_triton/common_utils.py

@@ -321,8 +321,6 @@ def get_config_filename(
     # NOTE(woosuk): The current naming convention uses w2.shape[2], which
     # is the intermediate size after silu_and_mul.
     N = shard_intermediate_size // 2
-    if use_int4_w4a16:
-        N = N // 2
 
     filename = get_config_file_name(
         num_experts,

benchmark/kernels/fused_moe_triton/tuning_fused_moe_triton.py

@@ -281,8 +281,6 @@ class BenchmarkWorker:
         block_n = block_shape[0] if block_shape else 0
         block_k = block_shape[1] if block_shape else 0
         N = shard_intermediate_size // 2
-        if use_int4_w4a16:
-            N = N // 2
         op_config = get_moe_configs(
             num_experts,
             N,

test/registered/unit/layers/moe/test_fused_moe_triton_config.py

@@ -1,4 +1,5 @@
 import json
+import sys
 from pathlib import Path
 
 from sglang.test.ci.ci_register import register_cpu_ci
@@ -9,6 +10,10 @@ register_cpu_ci(est_time=9, suite="base-a-test-cpu")
 from sglang.srt.layers.moe.moe_runner.triton_utils import fused_moe_triton_config
 from sglang.srt.runtime_context import get_context
 
+BENCHMARK_DIR = Path(__file__).parents[5] / "benchmark" / "kernels" / "fused_moe_triton"
+sys.path.insert(0, str(BENCHMARK_DIR))
+import common_utils  # noqa: E402
+
 
 def test_h200_bf16_config_is_available_for_current_triton_runtime():
     config_path = (
@@ -68,6 +73,30 @@ def test_down_moe_reuses_tuned_up_config_when_separate_config_is_absent(
         fused_moe_triton_config.get_moe_configs.cache_clear()
 
 
+def test_int4_tuner_filename_uses_runtime_down_projection_dimension(monkeypatch):
+    monkeypatch.setattr(
+        common_utils,
+        "get_config_file_name",
+        lambda E, N, *_args: f"E={E},N={N}.json",
+    )
+
+    filename = common_utils.get_config_filename(
+        num_experts=256,
+        shard_intermediate_size=512,
+        hidden_size=1024,
+        topk=8,
+        dtype=None,
+        use_fp8_w8a8=False,
+        use_int8_w8a8=False,
+        use_int8_w8a16=False,
+        use_int4_w4a16=True,
+        per_channel_quant=False,
+        block_shape=[128, 128],
+    )
+
+    assert filename == "E=256,N=256.json"
+
+
 if __name__ == "__main__":
     import sys