submission 565670

"""
MoE MXFP4 v71 — tile_k=128 for ALL shapes.

v67: tile_k=128 for E=257 only → 161us bench
Hypothesis: tile_k=128 may also help E=33 d=512 (K=512 → 4 iterations).
E=33 d=2048 keeps tile_k=256 (K=2048 → 8 iterations already fine).
"""
import os
os.environ["AITER_USE_NT"] = "1"

import sys
import functools
import torch
from task import input_t, output_t
from aiter import ActivationType, QuantType
from aiter.fused_moe import fused_moe
import aiter.fused_moe as _fmoe

# Block_m overrides (from v49)
_block_m_overrides = {
    (128, 9, 33, 512): 32,
    (128, 9, 33, 2048): 32,
    (512, 9, 33, 2048): 64,
}

@functools.lru_cache(maxsize=2048)
def _custom_get_block_size_M(token, topk, expert, inter_dim):
    key = (token, topk, expert, inter_dim)
    if key in _block_m_overrides:
        return _block_m_overrides[key]
    cu_num = _fmoe.get_cu_num()
    tileN = 128
    tgN = (inter_dim + tileN - 1) // tileN
    support_list = [32, 64, 128]
    tmp = []
    for el in support_list:
        max_num_tokens = token * topk + expert * el - topk
        tg_num = tgN * (max_num_tokens + el - 1) // el
        rnd = (tg_num + cu_num - 1) // cu_num
        empty = cu_num - tg_num % cu_num
        tmp.append((rnd, empty, el))
    return sorted(tmp, key=lambda x: x[:2])[0][-1]

_fmoe.get_block_size_M = _custom_get_block_size_M
try:
    _fmoe.get_2stage_cfgs.cache_clear()
except:
    pass


_flydsl_ready = False

def _flydsl_stage2_wrapper(inter_states, w1, w2, sorted_token_ids,
                            sorted_expert_ids, num_valid_ids, out, topk,
                            w2_scale=None, a2_scale=None, sorted_weights=None,
                            **_kwargs):
    from aiter.ops.flydsl.moe_kernels import flydsl_moe_stage2
    E = w2.shape[0]
    K = w2.shape[2] * 2  # FP4 packed, actual K = d_expert_pad
    # tile_k=128 for small/medium K, tile_k=256 for large K
    if K <= 512:
        tile_k = 128
    else:
        tile_k = 256

    flydsl_moe_stage2(
        inter_states=inter_states,
        w2=w2,
        sorted_token_ids=sorted_token_ids,
        sorted_expert_ids=sorted_expert_ids,
        num_valid_ids=num_valid_ids,
        out=out,
        topk=topk,
        tile_m=32,
        tile_n=128,
        tile_k=tile_k,
        a_dtype="fp4",
        b_dtype="fp4",
        out_dtype="bf16",
        mode="atomic",
        w2_scale=w2_scale,
        a2_scale=a2_scale,
        sorted_weights=sorted_weights,
    )


_orig_get_2stage_cfgs = _fmoe.get_2stage_cfgs

@functools.lru_cache(maxsize=2048)
def _patched_get_2stage_cfgs(*args, **kwargs):
    metadata = _orig_get_2stage_cfgs(*args, **kwargs)
    if _flydsl_ready and not metadata.run_1stage and metadata.stage2 is not None:
        metadata.stage2 = functools.partial(_flydsl_stage2_wrapper)
    return metadata

_fmoe.get_2stage_cfgs = _patched_get_2stage_cfgs


def _compile_flydsl():
    global _flydsl_ready
    try:
        from aiter.ops.flydsl.moe_kernels import _get_compiled_stage2
        import time
        print("Compiling flydsl stage2 (v71 tk=128 all small K)...", file=sys.stderr)

        # E=257 d=256: tile_k=128 (K=256)
        t0 = time.time()
        _get_compiled_stage2(
            model_dim=7168, inter_dim=256, experts=257, topk=9,
            tile_m=32, tile_n=128, tile_k=128,
            doweight=True, a_dtype="fp4", b_dtype="fp4",
            out_dtype="bf16", accumulate=True,
        )
        print(f"  E=257 d=256 tk=128: {time.time()-t0:.1f}s", file=sys.stderr)

        # E=33 d=512: tile_k=128 (K=512)
        t0 = time.time()
        _get_compiled_stage2(
            model_dim=7168, inter_dim=512, experts=33, topk=9,
            tile_m=32, tile_n=128, tile_k=128,
            doweight=True, a_dtype="fp4", b_dtype="fp4",
            out_dtype="bf16", accumulate=True,
        )
        print(f"  E=33 d=512 tk=128: {time.time()-t0:.1f}s", file=sys.stderr)

        # E=33 d=2048: tile_k=256 (K=2048, standard)
        t0 = time.time()
        _get_compiled_stage2(
            model_dim=7168, inter_dim=2048, experts=33, topk=9,
            tile_m=32, tile_n=128, tile_k=256,
            doweight=True, a_dtype="fp4", b_dtype="fp4",
            out_dtype="bf16", accumulate=True,
        )
        print(f"  E=33 d=2048 tk=256: {time.time()-t0:.1f}s", file=sys.stderr)

        _flydsl_ready = True
        print("flydsl ready (v71)", file=sys.stderr)
    except Exception as e:
        print(f"flydsl FAILED: {e}", file=sys.stderr)
        import traceback
        traceback.print_exc(file=sys.stderr)


_warmed = False
def _warmup():
    global _warmed
    if _warmed:
        return
    _warmed = True
    _compile_flydsl()
    configs = [
        (2, 256, 1, 7168, 256, 8),
        (2, 32, 1, 7168, 512, 8),
        (2, 32, 1, 7168, 2048, 8),
    ]
    for bs, n_routed, n_shared, d_hidden, d_expert, n_experts_per_token in configs:
        E = n_routed + n_shared
        total_topk = n_experts_per_token + n_shared
        d_hidden_pad = ((d_hidden + 255) // 256) * 256
        d_expert_pad = ((d_expert + 255) // 256) * 256
        h = torch.randn(bs, d_hidden, dtype=torch.bfloat16, device="cuda")
        w1 = torch.empty(E, 2 * d_expert_pad, d_hidden_pad // 2,
                         dtype=torch.float4_e2m1fn_x2, device="cuda")
        w2 = torch.empty(E, d_hidden_pad, d_expert_pad // 2,
                         dtype=torch.float4_e2m1fn_x2, device="cuda")
        w1_s = torch.empty(E, 2 * d_expert_pad, d_hidden_pad // 32,
                           dtype=torch.float8_e8m0fnu, device="cuda")
        w2_s = torch.empty(E, d_hidden_pad, d_expert_pad // 32,
                           dtype=torch.float8_e8m0fnu, device="cuda")
        topk_w = torch.ones(bs, total_topk, dtype=torch.float32, device="cuda")
        topk_i = torch.zeros(bs, total_topk, dtype=torch.int32, device="cuda")
        for t in range(bs):
            for k in range(n_experts_per_token):
                topk_i[t, k] = k % n_routed
            for k in range(n_shared):
                topk_i[t, n_experts_per_token + k] = n_routed + k
        try:
            fused_moe(h, w1, w2, topk_w, topk_i,
                      activation=ActivationType.Silu,
                      quant_type=QuantType.per_1x32,
                      w1_scale=w1_s, w2_scale=w2_s,
                      hidden_pad=d_hidden_pad - d_hidden,
                      intermediate_pad=d_expert_pad - d_expert)
            torch.cuda.synchronize()
            print(f"Warmup OK: E={E} d_e={d_expert}", file=sys.stderr)
        except Exception as e:
            print(f"Warmup FAIL: E={E} d_e={d_expert}: {e}", file=sys.stderr)
    print("Warmup complete (v71)", file=sys.stderr)

_warmup()


def custom_kernel(data: input_t) -> output_t:
    (hidden_states, gate_up_weight, down_weight,
     gate_up_weight_scale, down_weight_scale,
     gate_up_weight_shuffled, down_weight_shuffled,
     gate_up_weight_scale_shuffled, down_weight_scale_shuffled,
     topk_weights, topk_ids, config) = data

    return fused_moe(
        hidden_states,
        gate_up_weight_shuffled,
        down_weight_shuffled,
        topk_weights,
        topk_ids,
        activation=ActivationType.Silu,
        quant_type=QuantType.per_1x32,
        w1_scale=gate_up_weight_scale_shuffled,
        w2_scale=down_weight_scale_shuffled,
        hidden_pad=config["d_hidden_pad"] - config["d_hidden"],
        intermediate_pad=config["d_expert_pad"] - config["d_expert"],
    )