Fixed invalid warpSize in host code

[dongmin-ra]

Motivation

Fixed invalid warpSize in host(.cpp) code

Technical Details

Problem

• Right after this PR, GPU memory access fault occurred during internode EP execution.

Cause

• The cause is that this PR changed the logic to set warpSize to 64 only when __GFX8__ or __GFX9__ is defined, and otherwise to 32.
  • These macro variables are defined automatically by the compiler (amd clang++) when compiling device code.
  • However, when compiling host code (.cpp files), these macros are not defined.
• As a result, inside kernel code warpSize is set to 64, but in host code (e.g., dispatch_combine.cpp) it is set to 32.
• When launching the dispatch and combine kernels, the block dimension is set as warpSize * actualWarpNumPerBlock.
  • Since warpSize was incorrectly set to 32 in the host code, the block dimension ended up being half of the intended value, which caused the error.

Fix

• Explicitly define __GFX8__ or __GFX9__ in the CMake configuration based on the detected GPU architecture.

Test Plan

1. Apply the following changes to the examples/ops/dispatch_combine/test_dispatch_combine_internode.py file

diff --git a/examples/ops/dispatch_combine/test_dispatch_combine_internode.py b/examples/ops/dispatch_combine/test_dispatch_combine_internode.py
index 55d4ef2..bcf87e8 100644
--- a/examples/ops/dispatch_combine/test_dispatch_combine_internode.py
+++ b/examples/ops/dispatch_combine/test_dispatch_combine_internode.py
@@ -45,7 +45,7 @@ class EpDispatchCombineTestCase:
             num_experts_per_rank=16,
             # num_experts_per_rank=256 // world_size,
             num_experts_per_token=8,
-            warp_num_per_block=16,
+            warp_num_per_block=1,
             block_num=64,
             max_token_type_size=2,
             kernel_type=mori.ops.EpDispatchCombineKernelType.InterNode,

2. Execute the example script

export MORI_DISABLE_P2P=1
torchrun --local-ranks-filter 0 \
                --role rank \
                --nnodes=1 \
                --node_rank=0 \
                --nproc_per_node=1 \
                --master_addr=127.0.0.1 \
                --master_port=1234 \
                examples/ops/dispatch_combine/test_dispatch_combine_internode.py --max-tokens 16

Test Result

• Before modification : Memory acces fault error occurs.

Memory access fault by GPU node-3 (Agent handle: 0x820f5b0) on address 0x7efde4c00000. Reason: Unknown.
Memory access fault by GPU node-7 (Agent handle: 0xa625520) on address 0x7f6177600000. Reason: Unknown.
Memory access fault by GPU node-4 (Agent handle: 0x9df4740) on address 0x7f8066e00000. Reason: Unknown.
Memory access fault by GPU node-2 (Agent handle: 0x86d2ed0) on address 0x7ee328c00000. Reason: Unknown.
Memory access fault by GPU node-6 (Agent handle: 0x9f13da0) on address 0x7fae86a00000. Reason: Unknown.
Memory access fault by GPU node-8 (Agent handle: 0x9aa4c00) on address 0x7ef5c8c00000. Reason: Unknown.
Memory access fault by GPU node-9 (Agent handle: 0x976c110) on address 0x7f0b94c00000. Reason: Unknown.
Memory access fault by GPU node-5 (Agent handle: 0x8cb5700) on address 0x7fb4e7600000. Reason: Unknown.

• After modification : no error occurs.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[hnts03-moreh]

LGTM

[kyuhyeon-an]

👍