[Kernels] Clean up FusedMoeMethodBase and modular kernel setup. Remove extra arguments from modular kernel methods. (vllm-project#22035)

Signed-off-by: Bill Nell <[email protected]>
Co-authored-by: Michael Goin <[email protected]>

Author: 2 people

.buildkite/test-pipeline.yaml

@@ -399,6 +399,7 @@ steps:
 - label: Kernels MoE Test %N
   mirror_hardwares: [amdexperimental]
   source_file_dependencies:
+  - csrc/quantization/cutlass_w8a8/moe/
   - csrc/moe/
   - tests/kernels/moe
   - vllm/model_executor/layers/fused_moe/

docs/design/fused_moe_modular_kernel.md

@@ -175,11 +175,19 @@ implementations that input `FusedMoEActivationFormat.Standard` support chunking
 
 ### FusedMoEModularKernel Initialization
 
-`FusedMoEMethodBase` class has 2 methods that are collectively responsible in creating the `FusedMoEModularKernel` object. They are,
+`FusedMoEMethodBase` class has 3 methods that are collectively responsible in creating the `FusedMoEModularKernel` object. They are,
 
+* maybe_make_prepare_finalize,
 * select_gemm_impl, and
 * init_prepare_finalize
 
+#### maybe_make_prepare_finalize
+
+The `maybe_make_prepare_finalize` method is responsbile for constructing an instance of `FusedMoEPrepareAndFinalize` when appropriate based on the current all2all backend, e.g. when EP + DP is enabled.  The base class method currently constructs all the `FusedMoEPrepareAndFinalize` objects for the EP+DP case.  Derived classes can override this method to construct prepare/finalize objects for different scenarios, e.g. `ModelOptNvFp4FusedMoE` can construct a `FlashInferCutlassMoEPrepareAndFinalize` for the EP+TP case.
+Please refer to the implementations in,
+
+* `ModelOptNvFp4FusedMoE`
+
 #### select_gemm_impl
 
 The `select_gemm_impl` method is undefined in the base class. It is the responsibility of the derived class to implement a method that constructs a valid/appropriate `FusedMoEPermuteExpertsUnpermute` object.

examples/offline_inference/data_parallel.py

@@ -70,12 +70,27 @@ def parse_args():
         default=64,
         help=("Maximum number of sequences to be processed in a single iteration."),
     )
+    parser.add_argument(
+        "--max-model-len",
+        type=int,
+        help=("Maximum number of tokens to be processed in a single iteration."),
+    )
+    parser.add_argument(
+        "--timeout",
+        type=int,
+        default=300,
+        help=("Number of seconds before unresponsive process is killed."),
+    )
     parser.add_argument(
         "--gpu-memory-utilization",
         type=float,
         default=0.8,
         help=("Fraction of GPU memory vLLM is allowed to allocate (0.0, 1.0]."),
     )
+    parser.add_argument(
+        "--quantization",
+        type=str,
+    )
     return parser.parse_args()
 
 
@@ -90,7 +105,9 @@ def main(
     enforce_eager,
     trust_remote_code,
     max_num_seqs,
+    max_model_len,
     gpu_memory_utilization,
+    quantization,
 ):
     os.environ["VLLM_DP_RANK"] = str(global_dp_rank)
     os.environ["VLLM_DP_RANK_LOCAL"] = str(local_dp_rank)
@@ -142,7 +159,9 @@ def start(rank):
         enable_expert_parallel=True,
         trust_remote_code=trust_remote_code,
         max_num_seqs=max_num_seqs,
+        max_model_len=max_model_len,
         gpu_memory_utilization=gpu_memory_utilization,
+        quantization=quantization,
     )
     outputs = llm.generate(prompts, sampling_params)
     # Print the outputs.
@@ -198,14 +217,16 @@ def start(rank):
                 args.enforce_eager,
                 args.trust_remote_code,
                 args.max_num_seqs,
+                args.max_model_len,
                 args.gpu_memory_utilization,
+                args.quantization,
             ),
         )
         proc.start()
         procs.append(proc)
     exit_code = 0
     for proc in procs:
-        proc.join(timeout=300)
+        proc.join(timeout=args.timeout)
         if proc.exitcode is None:
             print(f"Killing process {proc.pid} that didn't stop within 5 minutes.")
             proc.kill()