pytorch - 💡(How to fix) Fix [Inductor] Block-wise quantization (MXFP8) fusion rejects FloorDiv broadcast index before reindex runs [1 comments, 1 participants]

#!/usr/bin/env python3
"""
Reproducer: Inductor fails to fuse block-wise quantization with preceding reduction.

Pattern: RMSNorm → block-wise MXFP8 quantization
  - RMSNorm: variance reduction over K (rnumel=K)
  - MXFP8 quant: block amax reduction over block_size=32 (rnumel=32),
    then pointwise fp8 conversion using broadcast scale (FloorDiv index)

Expected: 2 kernels (variance reduction + fused block_amax/scale/quantize)
Actual:   3 kernels (variance reduction, block_amax+scale_encode, fp8_quantize)

Root cause: fusable_read_and_write() in scheduler.py requires exact index
expression match between producer writes and consumer reads. The block amax
reduction writes with index `256*p0 + p1` (ranges p1∈[0,256)), but the fp8
quantize pointwise reads with `256*p0 + (p1//32)` (ranges p1∈[0,8192)) —
the FloorDiv broadcast pattern is not recognized as fusable.

Affects: Any block-wise quantization pattern (MXFP8, MXFP4, etc.) where
a per-block scale is broadcast back to element-level granularity.

Usage:
  python reproducer.py               # auto-detect device
  python reproducer.py --device cuda  # force CUDA
  python reproducer.py --device xpu   # force XPU
  python reproducer.py --device cpu   # CPU (for reference, no GPU needed)
  python reproducer.py --debug        # also dump TORCH_COMPILE_DEBUG
"""

import argparse
import os
import sys

import torch
import torch._inductor.metrics as metrics


# ---------------------------------------------------------------------------
# Pattern: RMSNorm + block-wise MXFP8 quantization
# Verbatim from vLLM MXFP8 quantization path — DO NOT simplify
# ---------------------------------------------------------------------------

MXFP8_BLOCK_SIZE = 32


def mxfp8_quantize(x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
    """Block-wise MXFP8 E4M3 quantization.

    Decomposed op sequence from real vLLM FX graph:
      to_fp32 → unflatten(K → num_blocks × 32) → abs → amax(dim=-1)
      → clamp(min=FLT_MIN) → log2 → floor → +127 → clamp(0,254) → to_uint8
      → exp2(scale−127) → unsqueeze(-1) → div → flatten → to_fp8e4m3
    """
    K = x.shape[-1]
    num_blocks = K // MXFP8_BLOCK_SIZE

    x_fp32 = x.to(torch.float32)
    x_blocked = x_fp32.unflatten(-1, (num_blocks, MXFP8_BLOCK_SIZE))

    # Per-block amax — reduction over block_size=32
    amax = x_blocked.abs().amax(dim=-1)
    amax = amax.clamp(min=torch.finfo(torch.float32).tiny)

    # Scale encoding (e8m0 format)
    scale_biased = torch.floor(torch.log2(amax)) + 127.0
    scale_biased = scale_biased.clamp(0, 254)
    scales_uint8 = scale_biased.to(torch.uint8)

    # Quantize: broadcast scale back to elements via unsqueeze
    descale = torch.exp2(scale_biased - 127.0)
    x_scaled = x_blocked / descale.unsqueeze(-1)

    # Final dtype conversions
    x_fp8 = x_scaled.flatten(-2).to(torch.float8_e4m3fn)
    scales = scales_uint8.view(torch.float8_e8m0fnu)
    return x_fp8, scales


def rms_norm_then_mxfp8_quant(
    x: torch.Tensor, norm_weight: torch.Tensor, eps: float = 1e-6
) -> tuple[torch.Tensor, torch.Tensor]:
    """RMSNorm followed by block-wise MXFP8 quantization."""
    x_fp32 = x.to(torch.float32)
    variance = x_fp32.pow(2).mean(dim=-1, keepdim=True)
    normed = x_fp32 * torch.rsqrt(variance + eps)
    normed = normed.to(torch.bfloat16) * norm_weight
    return mxfp8_quantize(normed)


# ---------------------------------------------------------------------------
# Standalone MXFP8 quantize (no RMSNorm) — isolates the block-broadcast gap
# ---------------------------------------------------------------------------


def mxfp8_quant_only(x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
    """Standalone block-wise MXFP8 quantization (no preceding norm)."""
    return mxfp8_quantize(x)


# ---------------------------------------------------------------------------
# Test harness
# ---------------------------------------------------------------------------


def detect_device() -> str:
    if torch.cuda.is_available():
        return "cuda"
    if hasattr(torch, "xpu") and torch.xpu.is_available():
        return "xpu"
    return "cpu"


def sync(device: str) -> None:
    if device == "cuda":
        torch.cuda.synchronize()
    elif device == "xpu":
        torch.xpu.synchronize()


def count_kernels(fn, *args, device: str) -> int:
    torch._dynamo.reset()
    metrics.reset()
    compiled = torch.compile(fn, fullgraph=True)
    compiled(*args)
    sync(device)
    return metrics.generated_kernel_count


def main() -> None:
    parser = argparse.ArgumentParser(description="Inductor block-broadcast fusion gap reproducer")
    parser.add_argument("--device", type=str, default=None, help="Device: cuda, xpu, or cpu")
    parser.add_argument("--debug", action="store_true", help="Enable TORCH_COMPILE_DEBUG")
    args = parser.parse_args()

    device = args.device or detect_device()
    if args.debug:
        os.environ["TORCH_COMPILE_DEBUG"] = "1"

    print(f"PyTorch version: {torch.__version__}")
    print(f"Device: {device}")
    print(f"float8_e4m3fn support: {hasattr(torch, 'float8_e4m3fn')}")
    print(f"float8_e8m0fnu support: {hasattr(torch, 'float8_e8m0fnu')}")
    print()

    K = 8192
    torch.manual_seed(42)

    print("=" * 70)
    print("Test 1: RMSNorm + MXFP8 block-wise quantization")
    print("=" * 70)
    print()
    print(f"  Pattern: variance_red(rnumel={K}) → norm_pw → block_amax_red(rnumel={MXFP8_BLOCK_SIZE})")
    print(f"           → scale_encode_pw → fp8_quantize_pw(reads scale via FloorDiv)")
    print()
    print(f"  Note: M=1 may produce 2 kernels (optimal) because Inductor collapses")
    print(f"  the M dimension, simplifying the iteration domain. The FloorDiv gap")
    print(f"  manifests at M≥8 where the block structure is fully preserved.")
    print()

    for M in [1, 2, 4, 8, 32, 64]:
        x = torch.randn(M, K, dtype=torch.bfloat16, device=device)
        w = torch.randn(K, dtype=torch.bfloat16, device=device)
        n = count_kernels(rms_norm_then_mxfp8_quant, x, w, device=device)
        status = "OK" if n <= 2 else "SUBOPTIMAL"
        print(f"  M={M:5d}  K={K}  kernels={n}  (optimal=2)  [{status}]")

    print()
    print("=" * 70)
    print("Test 2: Standalone MXFP8 block-wise quantization (no RMSNorm)")
    print("=" * 70)
    print()
    print(f"  Pattern: block_amax_red(rnumel={MXFP8_BLOCK_SIZE}) → scale_encode_pw")
    print(f"           → fp8_quantize_pw(reads scale via FloorDiv)")
    print()

    for M in [1, 2, 4, 8, 32, 64]:
        x = torch.randn(M, K, dtype=torch.bfloat16, device=device)
        n = count_kernels(mxfp8_quant_only, x, device=device)
        status = "OK" if n <= 1 else "SUBOPTIMAL"
        print(f"  M={M:5d}  K={K}  kernels={n}  (optimal=1)  [{status}]")

    print()
    print("=" * 70)
    print("Fusion log (set TORCH_LOGS='+fusion' to see rejection details)")
    print("=" * 70)
    print()
    print("Expected rejection messages:")
    print("  - 'numel/rnumel mismatch (reduce)' between variance and block_amax")
    print("  - 'memory deps did not match' between block_amax and fp8_quantize")
    print("    (due to FloorDiv broadcast index: write=256*p0+p1 vs read=256*p0+(p1//32))")
    print()
    print("To see full fusion trace:")
    print(f"  TORCH_LOGS='+fusion' python {sys.argv[0]}")


if __name__ == "__main__":
    main()

---

Collecting environment information...
PyTorch version: N/A
Is debug build: N/A
CUDA used to build PyTorch: N/A
ROCM used to build PyTorch: N/A

OS: Ubuntu 25.04 (x86_64)
GCC version: (Ubuntu 14.2.0-19ubuntu2) 14.2.0
Clang version: Could not collect
CMake version: version 3.31.6
Libc version: glibc-2.41

Python version: 3.13.12 | packaged by conda-forge | (main, Feb  5 2026, 05:53:46) [GCC 14.3.0] (64-bit runtime)
Python platform: Linux-6.14.0-37-generic-x86_64-with-glibc2.41
Is CUDA available: N/A
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: Could not collect
Nvidia driver version: Could not collect
cuDNN version: Could not collect
Is XPU available: N/A
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: N/A
Caching allocator config: N/A

CPU:
Architecture:                            x86_64
CPU op-mode(s):                          32-bit, 64-bit
Address sizes:                           46 bits physical, 48 bits virtual
Byte Order:                              Little Endian
CPU(s):                                  20
On-line CPU(s) list:                     0-19
Vendor ID:                               GenuineIntel
BIOS Vendor ID:                          Intel(R) Corporation
Model name:                              Intel(R) Core(TM) Ultra 7 265K
BIOS Model name:                         Intel(R) Core(TM) Ultra 7 265K To Be Filled By O.E.M. CPU @ 3.9GHz
BIOS CPU family:                         774
CPU family:                              6
Model:                                   198
Thread(s) per core:                      1
Core(s) per socket:                      20
Socket(s):                               1
Stepping:                                2
CPU(s) scaling MHz:                      44%
CPU max MHz:                             5500.0000
CPU min MHz:                             800.0000
BogoMIPS:                                7756.80
Flags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault ssbd ibrs ibpb stibp ibrs_enhanced tpr_shadow flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid rdt_a rdseed adx smap clflushopt clwb intel_pt sha_ni xsaveopt xsavec xgetbv1 xsaves split_lock_detect user_shstk avx_vnni lam wbnoinvd dtherm ida arat pln pts hwp hwp_notify hwp_act_window hwp_epp hwp_pkg_req hfi vnmi umip pku ospke waitpkg gfni vaes vpclmulqdq rdpid bus_lock_detect movdiri movdir64b fsrm md_clear serialize arch_lbr ibt flush_l1d arch_capabilities
Virtualization:                          VT-x
L1d cache:                               704 KiB (18 instances)
L1i cache:                               1.1 MiB (18 instances)
L2 cache:                                36 MiB (11 instances)
L3 cache:                                30 MiB (1 instance)
NUMA node(s):                            1
NUMA node0 CPU(s):                       0-19
Vulnerability Gather data sampling:      Not affected
Vulnerability Ghostwrite:                Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Not affected
Vulnerability Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Not affected
Vulnerability Spec rstack overflow:      Not affected
Vulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:                Mitigation; Enhanced / Automatic IBRS; IBPB conditional; PBRSB-eIBRS Not affected; BHI BHI_DIS_S
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Not affected
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Mitigation; IBPB before exit to userspace

Versions of relevant libraries:
[pip3] No relevant packages
[conda] No relevant packages