[XPU] [Bugfix] process ragged weights in xpu linear backend (#52118)
Signed-off-by: Zhu, Zufang <zufang.zhu@intel.com>
zofia committed
Aug 14, 2026 at 09:11 UTC
fe4c5dcd4cc71c2c1dc05375be497e413425aad7
1 file changed
+29
-1
vllm/model_executor/kernels/linear/scaled_mm/xpu.py
+29
-1
@@ -1,6 +1,7 @@
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
+import math
5
from collections.abc import Sequence
6
7
import torch
@@ -204,7 +205,34 @@ class XPUFp8BlockScaledMMKernel(Fp8BlockScaledMMLinearKernel):
205
)
206
scale = getattr(layer, scale_attr)
207
207
- # Checkpoint scale is [n_blocks, k_blocks] (one value per 128x128 tile).
208
+ # Ragged N (N % block_n != 0): oneDNN needs n_blocks to divide N.
209
+ # Weight untouched; only repeat scale rows to a finer N-group gn that
210
+ # divides both N and block_n (gn = gcd(N, block_n)):
211
+ # scale [ceil(N/block_n), K/block_k] --> [N/gn, K/block_k]
212
+ # oneDNN only accepts gn that is a multiple of 16, and gcd(N, block_n)
213
+ # is a power of two (block_n=128), so gn must be >= 16. No-op when
214
+ # N % block_n == 0.
215
+ block_n, block_k = self.weight_group_shape
216
+ N, K = layer.weight.shape
217
+ if N % block_n != 0:
218
+ gn = math.gcd(N, block_n)
219
+ assert gn % 16 == 0, (
220
+ f"XPU block-scaled FP8: N ({N}) yields group width {gn}, but "
221
+ f"oneDNN only supports multiples of 16; this weight shape is "
222
+ f"unsupported."
223
+ )
224
+ col_start = torch.arange(N // gn, device=scale.device) * gn
225
+ src_idx = torch.div(col_start, block_n, rounding_mode="floor")
226
+ scale = scale.index_select(0, src_idx).contiguous()
227
+
228
+ # Ragged K needs the runtime activation scale expanded too, which we
229
+ # don't handle; DeepSeek/GLM keep K block-aligned, so fail loudly.
230
+ assert K % block_k == 0, (
231
+ f"XPU block-scaled FP8 requires K ({K}) to be a multiple of the "
232
+ f"weight block size ({block_k}); ragged-K weights are unsupported."
233
+ )
234
+
235
+ # Checkpoint scale is [n_blocks, k_blocks] (one value per block tile).
236
# oneDNN fp8_gemm requires contiguous [k_blocks, n_blocks] layout.
237
# We store the transposed contiguous buffer as a .t() view so that:
238
# - MLA's scaled_dequantize still sees [n_blocks, k_blocks] shape