feat: implement sm90 megamoe phase3 l1 wgmma

megamoe_dev_test_scripts/phase3/l1_wgmma_single_tile.py

@@ -0,0 +1,169 @@
+import argparse
+import inspect
+import os
+import pathlib
+import random
+import sys
+from typing import 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 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', '8362'))
+
+    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 interleaved_to_natural_n(n: torch.Tensor, half_n: int, gran: int = 8) -> torch.Tensor:
+    pair_group = n // (2 * gran)
+    offset = n - pair_group * (2 * gran)
+    gate_n = pair_group * gran + offset
+    up_n = half_n + pair_group * gran + offset - gran
+    return torch.where(offset < gran, gate_n, up_n)
+
+
+def make_weights(num_experts_per_rank: int, hidden: int, intermediate_hidden: int):
+    torch.manual_seed(2027 + dist.get_rank())
+    l1_weights = (torch.randn(
+        (num_experts_per_rank, intermediate_hidden * 2, hidden),
+        dtype=torch.float32, device='cuda') * 0.5).to(torch.float8_e4m3fn)
+    l2_weights = (torch.randn(
+        (num_experts_per_rank, hidden, intermediate_hidden),
+        dtype=torch.float32, device='cuda') * 0.5).to(torch.float8_e4m3fn)
+
+    l1_weights_sf = torch.empty(
+        (num_experts_per_rank, ceil_div(intermediate_hidden * 2, 128), hidden // 128),
+        dtype=torch.float32, device='cuda')
+    for expert in range(num_experts_per_rank):
+        for n_group in range(l1_weights_sf.shape[1]):
+            for k_group in range(l1_weights_sf.shape[2]):
+                l1_weights_sf[expert, n_group, k_group] = 0.75 + 0.125 * n_group + 0.0625 * k_group
+
+    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 reference_l1_accum(buffer: deep_gemm.SymmBuffer,
+                       l1_weights: torch.Tensor,
+                       l1_weights_sf: torch.Tensor,
+                       hidden: int,
+                       intermediate_hidden: int) -> torch.Tensor:
+    block_m = 128
+    block_n = 128
+    x_fp8 = buffer.l1_acts[:block_m]
+    x_sf = buffer.l1_acts_sf[:block_m, :hidden // 128]
+    w_fp8 = l1_weights[0, :block_n]
+
+    n = torch.arange(block_n, device='cuda')
+    natural_n = interleaved_to_natural_n(n, intermediate_hidden)
+    n_groups = natural_n // 128
+
+    expected = torch.zeros((block_m, block_n), dtype=torch.float32, device='cuda')
+    for k_group in range(hidden // 128):
+        k_start = k_group * 128
+        k_end = k_start + 128
+        x = x_fp8[:, k_start:k_end].to(torch.float32)
+        w = w_fp8[:, k_start:k_end].to(torch.float32)
+        sfa = x_sf[:, k_group].to(torch.float32)
+        sfb = l1_weights_sf[0, n_groups, k_group].to(torch.float32)
+        expected += (x @ w.t()) * sfa[:, None] * sfb[None, :]
+    return expected
+
+
+def run_case(args: argparse.Namespace, group: dist.ProcessGroup, rank_idx: int, num_ranks: int) -> None:
+    hidden = args.hidden
+    intermediate_hidden = args.intermediate_hidden
+    num_tokens = args.num_tokens
+    num_topk = 1
+    num_experts = args.num_experts if args.num_experts is not None else num_ranks
+    num_experts_per_rank = num_experts // num_ranks
+    assert num_experts % num_ranks == 0
+    assert num_experts_per_rank >= 1
+
+    buffer = deep_gemm.get_symm_buffer_for_mega_moe(
+        group, num_experts, args.num_max_tokens_per_rank, num_topk,
+        hidden, intermediate_hidden)
+    weights = make_weights(num_experts_per_rank, hidden, intermediate_hidden)
+
+    torch.manual_seed(1234 + rank_idx)
+    x = (torch.randn((num_tokens, hidden), dtype=torch.float32, device='cuda') * 0.5).to(torch.float8_e4m3fn)
+    x_sf = torch.rand((num_tokens, hidden // 128), dtype=torch.float32, device='cuda') + 0.75
+    local_global_expert = rank_idx * num_experts_per_rank
+    topk_idx = torch.full((num_tokens, num_topk), local_global_expert, dtype=torch.long, device='cuda')
+    topk_weights = torch.ones((num_tokens, num_topk), dtype=torch.float32, device='cuda')
+
+    buffer.x[:num_tokens].copy_(x)
+    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)
+
+    y = torch.empty((num_tokens, hidden), dtype=torch.bfloat16, device='cuda')
+    deep_gemm.fp8_mega_moe(y, weights[0], weights[1], buffer)
+    torch.cuda.synchronize()
+
+    expected = reference_l1_accum(buffer, weights[0][0], weights[0][1], hidden, intermediate_hidden)
+    actual = buffer.l1_accum_debug
+    torch.testing.assert_close(actual.cpu(), expected.cpu(), rtol=args.rtol, atol=args.atol)
+
+    dist.barrier(group=group)
+    if rank_idx == 0:
+        max_diff = (actual - expected).abs().max().item()
+        print(f'[PASSED] Phase 3 L1 WGMMA single tile, max_diff={max_diff:.6f}', flush=True)
+    buffer.destroy()
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description='SM90 MegaMoE Phase 3 L1 WGMMA single-tile correctness')
+    parser.add_argument('--num-tokens', type=int, default=128)
+    parser.add_argument('--num-max-tokens-per-rank', type=int, default=384)
+    parser.add_argument('--hidden', type=int, default=256)
+    parser.add_argument('--intermediate-hidden', type=int, default=128)
+    parser.add_argument('--num-experts', type=int, default=None)
+    parser.add_argument('--local-rank', type=int, default=None)
+    parser.add_argument('--atol', type=float, default=1e-2)
+    parser.add_argument('--rtol', type=float, default=5e-2)
+    args = parser.parse_args()
+
+    rank_idx, num_ranks, group = init_test_dist(args.local_rank)
+    assert torch.cuda.get_device_capability(torch.cuda.current_device())[0] == 9
+    assert args.num_tokens == 128
+    assert args.hidden % 128 == 0 and args.intermediate_hidden % 128 == 0
+
+    random.seed(5678 + rank_idx)
+    run_case(args, group, rank_idx, num_ranks)
+    dist.destroy_process_group()
+
+
+if __name__ == '__main__':
+    main()