[ROCm] Fix MoE kernel test failures on gfx950 (#37833)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Signed-off-by: Matthew Wong <Matthew.Wong2@amd.com>
Co-authored-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
Co-authored-by: Matthew Wong <Matthew.Wong2@amd.com>
This commit is contained in:
Andreas Karatzas
2026-03-25 13:46:40 -05:00
committed by GitHub
co-authored by Woosuk Kwon Matthew Wong
parent e38817fadb
commit 7d6917bef5
12 changed files with 478 additions and 86 deletions
@@ -741,11 +741,14 @@ class QuarkOCP_MX_MoEMethod(QuarkMoEMethod):
# TP=4 yields intermediate_size_per_partition=384), AITER raises:
# "device_gemm ... does not support this GEMM problem".
# Fall back to emulation in that case.
# For gpt_oss models, create_weights rounds up the dimensions
# internally, so the alignment check is skipped.
if (
not self.emulate
and self.use_rocm_aiter_moe
and self.ocp_mx_scheme is not None
and self.ocp_mx_scheme.startswith("w_mxfp4")
and self.model_type != "gpt_oss"
and moe.intermediate_size_per_partition % CK_MXFP4_MOE_DIM_ALIGNMENT != 0
):
logger.warning_once(
@@ -819,6 +822,18 @@ class QuarkOCP_MX_MoEMethod(QuarkMoEMethod):
"unpadded_hidden_size", hidden_size
)
# On GFX950, the GFX950MXScaleLayout swizzle requires
# hidden_size to be a multiple of 256 (SCALE_K = hidden_size / 32
# must be divisible by 8). Pad hidden_size for weight/scale
# allocation; the original value is preserved in unpadded_hidden_size.
# Only applies to the native (non-emulated) CK path on GFX950.
if (
self.model_type == "gpt_oss"
and current_platform.is_rocm()
and not self.emulate
):
hidden_size = round_up(hidden_size, 256)
# WEIGHTS
w13_weight = torch.nn.Parameter(
torch.empty(
@@ -615,8 +615,8 @@ def _per_token_group_quant_fp8(
# Avoid to divide zero
eps,
# Information for float8
fp8_min,
fp8_max,
fp8_min: tl.constexpr,
fp8_max: tl.constexpr,
use_ue8m0: tl.constexpr,
# Meta-parameters
BLOCK: tl.constexpr,
@@ -647,8 +647,12 @@ def _per_token_group_quant_fp8(
y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32)
# Quant
# Use multiply-by-reciprocal instead of division to match PyTorch's
# tensor/scalar division precision (GPU fast-division for constexpr
# divisors can introduce 1-ULP error that flips FP8 quantization at
# representable-value boundaries).
_absmax = tl.maximum(tl.max(tl.abs(y)), eps)
scale_raw = _absmax / fp8_max
scale_raw = _absmax * (1.0 / fp8_max)
y_s = tl.math.exp2(tl.ceil(tl.log2(scale_raw))) if use_ue8m0 else scale_raw
y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
@@ -667,8 +671,8 @@ def _silu_mul_per_token_group_quant_fp8_colmajor(
y_s_col_stride: tl.int64,
# Information for float8
eps,
fp8_min,
fp8_max,
fp8_min: tl.constexpr,
fp8_max: tl.constexpr,
use_ue8m0: tl.constexpr,
# Meta-parameters
GROUP_SIZE: tl.constexpr,
@@ -709,7 +713,7 @@ def _silu_mul_per_token_group_quant_fp8_colmajor(
# quant
_absmax = tl.maximum(tl.max(tl.abs(y), axis=1), eps)
scale_raw = _absmax / fp8_max
scale_raw = _absmax * (1.0 / fp8_max)
y_s = tl.math.exp2(tl.ceil(tl.log2(scale_raw))) if use_ue8m0 else scale_raw
y_s = tl.reshape(y_s, (BLOCK_M, 1))
y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
@@ -808,8 +812,8 @@ def _per_token_group_quant_fp8_colmajor(
# Avoid to divide zero
eps,
# Information for float8
fp8_min,
fp8_max,
fp8_min: tl.constexpr,
fp8_max: tl.constexpr,
use_ue8m0: tl.constexpr,
# Meta-parameters
BLOCK: tl.constexpr,
@@ -849,7 +853,7 @@ def _per_token_group_quant_fp8_colmajor(
y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32)
# Quant
_absmax = tl.maximum(tl.max(tl.abs(y)), eps)
scale_raw = _absmax / fp8_max
scale_raw = _absmax * (1.0 / fp8_max)
y_s = tl.math.exp2(tl.ceil(tl.log2(scale_raw))) if use_ue8m0 else scale_raw
y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)