[Grok-1] 1. upload moe configuration file for moe kernel optimization… (#193)

* [Grok-1] 1. upload moe configuration file for moe kernel optimization 2. support "--num-scheduler-steps" in benchmark_latency.py

* [Grok-1] 1. upload moe configuration file for moe kernel optimization 2. add copy of benchmark_latency.py to support "--num-scheduler-steps"

* [Grok-1] add option num-scheduler-steps in benchmark_latency.py

Author: kkHuang-amd

benchmarks/benchmark_latency.py

@@ -47,6 +47,7 @@ def main(args: argparse.Namespace):
         distributed_executor_backend=args.distributed_executor_backend,
         otlp_traces_endpoint=args.otlp_traces_endpoint,
         enable_prefix_caching=args.enable_prefix_caching,
+        num_scheduler_steps=args.num_scheduler_steps,
     )
 
     sampling_params = SamplingParams(
@@ -279,5 +280,10 @@ def run_to_completion(profile_dir: Optional[str] = None):
         type=str,
         default=None,
         help='Target URL to which OpenTelemetry traces will be sent.')
+    parser.add_argument(
+        "--num-scheduler-steps",
+        type=int,
+        default=1,
+        help="Maximum number of forward steps per scheduler call.")
     args = parser.parse_args()
     main(args)

vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X_OAM,dtype=fp8_w8a8.json

@@ -1,13 +1,13 @@
 {
     "8": {
         "BLOCK_SIZE_M": 16,
-        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_N": 64,
         "BLOCK_SIZE_K": 256,
         "GROUP_SIZE_M": 1,
-        "num_warps": 8,
+        "num_warps": 4,
         "num_stages": 0,
-        "waves_per_eu": 0,
+        "waves_per_eu": 1,
         "matrix_instr_nonkdim": 16,
-        "kpack": 2
+        "kpack": 1
     }
 }