Skip to content
KernelIndex
Search⌘K
Sign in

submission 526320

Hamza · python · License unknown
Kernel source · 31 lines ↓

Use it

Vendorable · source mirrored · license unknownView source →

No package. Vendor the mirrored source: 31 lines, June 9 Researcher Reciprocity License v1.0.

submission.py
curl "https://kernelindex.com/api/v1/implementations/kernelbot-amd-mxfp4-mm-526320?include=source"
interfacepython
Compatibility
measured onAMD Instinct MI355X
declared hardwareAMD Instinct MI355X
architecturesgfx950
dtypesbf16, mxfp4

Benchmark evidence

1 measurement across 1 GPU, fastest first.

Operation / workload
Hardware
Latency
Rank
Observed
AMD MXFP4 GEMMsuite of 6 cases
AMD Instinct MI355X
14.9µs
#542 of 1143
2026-03-10
Run →

Reported · How evidence levels are derived →

Source and license

sourceavailable
revision digestsha256:576fcde32bf60ecbd7fba49e04d3f8861d7d2c9cebc433d5811ff98e65619897
license declaredunknown
license concludedunknown
authorsHamza
imported2026-08-15

Techniques

Extracted from the mirrored source by pattern, never inferred. Each row cites its line.

fp4FP4 quant + FP4 GEMM reference: bf16 A, MXFP4 B -> MXFP4 per-1x32 quant A -> gemm_a4w4 -> bf16 C.

Kernel source

submission.py31 lines
"""
FP4 quant + FP4 GEMM reference: bf16 A, MXFP4 B -> MXFP4 per-1x32 quant A -> gemm_a4w4 -> bf16 C.
Quant logic follows aiter op_tests/test_gemm_a4w4.py (get_triton_quant(QuantType.per_1x32)).
"""
import aiter
from aiter import QuantType, dtypes

from task import input_t, output_t

QUANT_FUNC = aiter.get_triton_quant(QuantType.per_1x32)
GEMM_A4W4 = aiter.gemm_a4w4
OUTPUT_DTYPE = dtypes.bf16


def custom_kernel(data: input_t) -> output_t:
    """
    Reference: MXFP4 per-1x32 quant on A; B_shuffle, B_scale_sh from generate_input.
    gemm_a4w4 with bpreshuffle=True.
    """
    A, _, _, B_shuffle, B_scale_sh = data
    A_q, A_scale_sh = QUANT_FUNC(A, shuffle=True)

    return GEMM_A4W4(
        A_q,
        B_shuffle,
        A_scale_sh,
        B_scale_sh,
        dtype=OUTPUT_DTYPE,
        bpreshuffle=True,
    )
scrolls · 31 lines total

Source code from GPU Mode and the KernelBot dataset · June 9 Researcher Reciprocity License v1.0

Changes from previous submission

Against this author's previous submission submission 526200.

  ⋯ 9 unchanged lines
  QUANT_FUNC = aiter.get_triton_quant(QuantType.per_1x32)
  GEMM_A4W4 = aiter.gemm_a4w4
  OUTPUT_DTYPE = dtypes.bf16
- _CACHED_A = None
- _CACHED_A_Q = None
- _CACHED_A_SCALE_SH = None
   
   
- def _quantize_a_cached(A):
-     global _CACHED_A, _CACHED_A_Q, _CACHED_A_SCALE_SH
-  
-     # Benchmark mode reuses the same tensor object for timed repeats.
-     if A is _CACHED_A:
-         return _CACHED_A_Q, _CACHED_A_SCALE_SH
-  
-     A_q, A_scale_sh = QUANT_FUNC(A, shuffle=True)
-     _CACHED_A = A
-     _CACHED_A_Q = A_q
-     _CACHED_A_SCALE_SH = A_scale_sh
-     return A_q, A_scale_sh
-  
-  
  def custom_kernel(data: input_t) -> output_t:
      """
      Reference: MXFP4 per-1x32 quant on A; B_shuffle, B_scale_sh from generate_input.
      gemm_a4w4 with bpreshuffle=True.
      """
      A, _, _, B_shuffle, B_scale_sh = data
+     A_q, A_scale_sh = QUANT_FUNC(A, shuffle=True)
   
-     A_q, A_scale_sh = _quantize_a_cached(A)
      return GEMM_A4W4(
          A_q,
          B_shuffle,
scrolls · 35 diff lines total

Best evidence level for this revision: reported

JSON