chenchenghao/DeepGEMM
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: implement sm90 megamoe phase2 dispatch-only

Browse Source
This commit is contained in:
Xinyi Liu
2026-06-18 00:00:36 +08:00
parent 74eb59dfaa
commit 540e5aeadc
4 changed files with 622 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

266
megamoe_dev_test_scripts/phase2/dispatch_only_correctness.py Normal file
View File

@@ -0,0 +1,266 @@
import argparse
import inspect
import os
import pathlib
import random
import sys
from typing import List, Tuple
import torch
import torch.distributed as dist
REPO_ROOT = pathlib.Path(__file__).resolve().parents[2]
if str(REPO_ROOT) not in sys.path:
sys.path.insert(0, str(REPO_ROOT))
import deep_gemm
from deep_gemm.utils.math import ceil_div
def parse_tokens_list(value: str) -> List[int]:
return [int(item) for item in value.split(',') if item]
def gather_same_shape(tensor: torch.Tensor, group: dist.ProcessGroup) -> torch.Tensor:
gathered = [torch.empty_like(tensor) for _ in range(dist.get_world_size(group))]
dist.all_gather(gathered, tensor.contiguous(), group=group)
return torch.stack(gathered, dim=0)
def init_test_dist(local_rank_arg: int = None) -> Tuple[int, int, dist.ProcessGroup]:
local_rank = local_rank_arg if local_rank_arg is not None else int(os.environ.get('LOCAL_RANK', '0'))
rank = int(os.environ.get('RANK', '0'))
world_size = int(os.environ.get('WORLD_SIZE', '1'))
master_addr = os.environ.get('MASTER_ADDR', '127.0.0.1')
master_port = int(os.environ.get('MASTER_PORT', '8361'))
torch.cuda.set_device(local_rank)
sig = inspect.signature(dist.init_process_group)
params = {
'backend': 'nccl',
'init_method': f'tcp://{master_addr}:{master_port}',
'world_size': world_size,
'rank': rank,
}
if 'device_id' in sig.parameters:
params['device_id'] = torch.device(f'cuda:{local_rank}')
dist.init_process_group(**params)
torch.set_default_device('cuda')
return rank, world_size, dist.new_group(list(range(world_size)))
def get_block_m(num_tokens: int, num_ranks: int, num_topk: int, num_experts: int) -> int:
expected = num_tokens * num_ranks * num_topk / num_experts
if expected <= 16.5:
return 32
if expected <= 64.5:
return 64
return 128
def make_topk(num_tokens: int, num_experts: int, num_topk: int, rank_idx: int) -> Tuple[torch.Tensor, torch.Tensor]:
token_idx = torch.arange(num_tokens, device='cuda', dtype=torch.long).unsqueeze(1)
topk_slot = torch.arange(num_topk, device='cuda', dtype=torch.long).unsqueeze(0)
topk_idx = (token_idx * num_topk + topk_slot + rank_idx * 3) % num_experts
topk_weights = (token_idx.float() * 0.125 + topk_slot.float() * 0.25 + rank_idx + 1.0).contiguous()
return topk_idx.contiguous(), topk_weights.contiguous()
def make_weights(num_experts_per_rank: int, hidden: int, intermediate_hidden: int):
l1_weights = torch.randn(
(num_experts_per_rank, intermediate_hidden * 2, hidden),
dtype=torch.float32, device='cuda').to(torch.float8_e4m3fn)
l2_weights = torch.randn(
(num_experts_per_rank, hidden, intermediate_hidden),
dtype=torch.float32, device='cuda').to(torch.float8_e4m3fn)
l1_weights_sf = torch.ones(
(num_experts_per_rank, ceil_div(intermediate_hidden * 2, 128), hidden // 128),
dtype=torch.float32, device='cuda')
l2_weights_sf = torch.ones(
(num_experts_per_rank, ceil_div(hidden, 128), intermediate_hidden // 128),
dtype=torch.float32, device='cuda')
return deep_gemm.transform_weights_for_mega_moe(
(l1_weights, l1_weights_sf), (l2_weights, l2_weights_sf))
def build_expected_entries(all_topk_idx: torch.Tensor,
rank_idx: int,
num_experts_per_rank: int) -> Tuple[List[List[Tuple[int, int, int]]], torch.Tensor]:
num_ranks, num_tokens, num_topk = all_topk_idx.shape
local_start = rank_idx * num_experts_per_rank
local_end = local_start + num_experts_per_rank
entries: List[List[Tuple[int, int, int]]] = [[] for _ in range(num_experts_per_rank)]
counts = [0 for _ in range(num_experts_per_rank)]
all_topk_idx_cpu = all_topk_idx.cpu()
for src_rank in range(num_ranks):
for token_idx in range(num_tokens):
for topk_idx in range(num_topk):
expert_idx = int(all_topk_idx_cpu[src_rank, token_idx, topk_idx])
if local_start <= expert_idx < local_end:
local_expert = expert_idx - local_start
entries[local_expert].append((src_rank, token_idx, topk_idx))
counts[local_expert] += 1
return entries, torch.tensor(counts, dtype=torch.int64, device='cuda')
def verify_case(buffer: deep_gemm.SymmBuffer,
all_x_f32: torch.Tensor,
all_x_sf: torch.Tensor,
all_topk_idx: torch.Tensor,
all_topk_weights: torch.Tensor,
cumulative_stats: torch.Tensor,
rank_idx: int,
num_ranks: int,
num_experts: int,
num_topk: int,
block_m: int) -> None:
del num_topk
num_experts_per_rank = num_experts // num_ranks
sf_block_m = ((block_m + 127) // 128) * 128
num_sms = torch.cuda.get_device_properties(torch.cuda.current_device()).multi_processor_count
entries_by_expert, counts = build_expected_entries(all_topk_idx, rank_idx, num_experts_per_rank)
torch.testing.assert_close(cumulative_stats.cpu(), counts.to(torch.int32).cpu(), rtol=0, atol=0)
expert_status = buffer.expert_recv_count_sum[:num_experts_per_rank].detach().cpu().tolist()
arrival = buffer.l1_arrival_count.detach().cpu()
metadata = buffer.token_src_metadata.detach().cpu()
pool_block_offset = 0
local_start = rank_idx * num_experts_per_rank
for local_expert, expected_entries in enumerate(entries_by_expert):
count = len(expected_entries)
status = int(expert_status[local_expert])
assert (status & 0xffffffff) == count, (local_expert, status, count)
assert (status >> 32) == num_sms * num_ranks, (local_expert, status, num_sms, num_ranks)
expected_set = set(expected_entries)
actual_set = set()
base_pool_token = pool_block_offset * block_m
expected_global_expert = local_start + local_expert
for token_in_expert in range(count):
pool_token_idx = base_pool_token + token_in_expert
src_rank, src_token_idx, src_topk_idx = [int(v) for v in metadata[pool_token_idx].tolist()]
actual_entry = (src_rank, src_token_idx, src_topk_idx)
assert actual_entry in expected_set, (local_expert, token_in_expert, actual_entry, expected_set)
assert actual_entry not in actual_set, (local_expert, token_in_expert, actual_entry)
actual_set.add(actual_entry)
expert_idx = int(all_topk_idx[src_rank, src_token_idx, src_topk_idx].item())
assert expert_idx == expected_global_expert, (pool_token_idx, expert_idx, expected_global_expert)
torch.testing.assert_close(
buffer.l1_acts[pool_token_idx].to(torch.float32).cpu(),
all_x_f32[src_rank, src_token_idx].cpu(),
rtol=0, atol=0)
torch.testing.assert_close(
buffer.l1_topk_weights[pool_token_idx].cpu(),
all_topk_weights[src_rank, src_token_idx, src_topk_idx].cpu(),
rtol=0, atol=0)
sf_pool_token_idx = (pool_block_offset + token_in_expert // block_m) * sf_block_m + token_in_expert % block_m
torch.testing.assert_close(
buffer.l1_acts_sf[sf_pool_token_idx].cpu(),
all_x_sf[src_rank, src_token_idx].cpu(),
rtol=0, atol=0)
assert actual_set == expected_set, (local_expert, actual_set, expected_set)
for block_idx in range(ceil_div(count, block_m)):
expected_arrivals = min(block_m, count - block_idx * block_m)
actual_arrivals = int(arrival[pool_block_offset + block_idx].item())
assert actual_arrivals == expected_arrivals, (local_expert, block_idx, actual_arrivals, expected_arrivals)
pool_block_offset += ceil_div(count, block_m)
def run_case(args: argparse.Namespace,
group: dist.ProcessGroup,
rank_idx: int,
num_ranks: int,
buffer: deep_gemm.SymmBuffer,
weights,
num_tokens: int) -> None:
hidden = args.hidden
num_topk = args.num_topk
num_experts = args.num_experts
x_f32 = torch.randn((num_tokens, hidden), dtype=torch.float32, device='cuda')
x_fp8 = x_f32.to(torch.float8_e4m3fn)
x_sf = torch.rand((num_tokens, hidden // 128), dtype=torch.float32, device='cuda') + 0.5
topk_idx, topk_weights = make_topk(num_tokens, num_experts, num_topk, rank_idx)
all_x_f32 = gather_same_shape(x_fp8.to(torch.float32), group)
all_x_sf = gather_same_shape(x_sf, group)
all_topk_idx = gather_same_shape(topk_idx, group)
all_topk_weights = gather_same_shape(topk_weights, group)
buffer.x[:num_tokens].copy_(x_fp8)
buffer.x_sf[:num_tokens].copy_(x_sf)
buffer.topk_idx[:num_tokens].copy_(topk_idx)
buffer.topk_weights[:num_tokens].copy_(topk_weights)
torch.cuda.synchronize()
dist.barrier(group=group)
cumulative_stats = torch.zeros((num_experts // num_ranks,), dtype=torch.int32, device='cuda')
y = torch.empty((num_tokens, hidden), dtype=torch.bfloat16, device='cuda')
deep_gemm.fp8_mega_moe(y, weights[0], weights[1], buffer,
cumulative_local_expert_recv_stats=cumulative_stats)
torch.cuda.synchronize()
block_m = get_block_m(num_tokens, num_ranks, num_topk, num_experts)
verify_case(buffer, all_x_f32, all_x_sf, all_topk_idx, all_topk_weights,
cumulative_stats, rank_idx, num_ranks, num_experts, num_topk, block_m)
dist.barrier(group=group)
if rank_idx == 0:
print(f'[PASSED] tokens={num_tokens}, block_m={block_m}', flush=True)
def main() -> None:
parser = argparse.ArgumentParser(description='SM90 MegaMoE Phase 2 dispatch-only correctness')
parser.add_argument('--tokens-list', type=str, default='0,8,48,192')
parser.add_argument('--num-max-tokens-per-rank', type=int, default=384)
parser.add_argument('--hidden', type=int, default=512)
parser.add_argument('--intermediate-hidden', type=int, default=256)
parser.add_argument('--num-experts', type=int, default=16)
parser.add_argument('--num-topk', type=int, default=6)
parser.add_argument('--local-rank', type=int, default=None)
args = parser.parse_args()
local_rank = args.local_rank if args.local_rank is not None else int(os.environ.get('LOCAL_RANK', '0'))
rank_idx, num_ranks, group = init_test_dist(local_rank)
assert torch.cuda.get_device_capability(torch.cuda.current_device())[0] == 9
assert args.num_experts % num_ranks == 0
assert args.hidden % 128 == 0 and args.intermediate_hidden % 128 == 0
assert args.num_topk <= args.num_experts
torch.manual_seed(1234 + rank_idx)
random.seed(1234 + rank_idx)
tokens_list = parse_tokens_list(args.tokens_list)
max_tokens = max([args.num_max_tokens_per_rank] + tokens_list)
buffer = deep_gemm.get_symm_buffer_for_mega_moe(
group, args.num_experts, max_tokens, args.num_topk,
args.hidden, args.intermediate_hidden)
weights = make_weights(args.num_experts // num_ranks, args.hidden, args.intermediate_hidden)
if rank_idx == 0:
print(f'[Phase 2] ranks={num_ranks}, tokens_list={tokens_list}, '
f'hidden={args.hidden}, intermediate={args.intermediate_hidden}, '
f'experts={args.num_experts}, topk={args.num_topk}', flush=True)
for num_tokens in tokens_list:
run_case(args, group, rank_idx, num_ranks, buffer, weights, num_tokens)
dist.barrier(group=group)
buffer.destroy()
dist.destroy_process_group()
if rank_idx == 0:
print('[PASSED] SM90 MegaMoE Phase 2 dispatch-only correctness', flush=True)
if __name__ == '__main__':
main()