[XPU] Route block-quantized FP8 weights to the W8A8 kernel (#50787)
Signed-off-by: Chaojun Zhang <chaojun.zhang@intel.com>
Chaojun Zhang committed
Aug 12, 2026 at 16:57 UTC
8c011da6d0a45f2c9316ea3c658da754f770a178
2 files changed
+11
-1
.buildkite/intel_jobs/test-intel.yaml
+2
-1
@@ -196,7 +196,8 @@ steps:
196
- >-
197
bash .buildkite/scripts/hardware_ci/run-intel-test.sh
198
'cd tests &&
199
- pytest -v -s quantization/test_compressed_tensors.py::test_compressed_tensors_fp8'
199
+ pytest -v -s quantization/test_compressed_tensors.py::test_compressed_tensors_fp8 &&
200
+ pytest -v -s quantization/test_compressed_tensors.py::test_compressed_tensors_fp8_block_enabled'
201
- label: "XPU Graph"
202
depends_on:
203
- image-build-xpu
vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py
+9
@@ -817,6 +817,15 @@ class CompressedTensorsConfig(QuantizationConfig):
817
is_fp8_w8a8_supported = config is not None and (
818
config.kernel_config.linear_backend in ("xpu", "torch")
819
)
820
+ weight_quant_is_block_strategy = (
821
+ weight_quant
822
+ and weight_quant.strategy == QuantizationStrategy.BLOCK
823
+ )
824
+ if weight_quant_is_block_strategy and not is_fp8_w8a8_supported:
825
+ # On XPU, block quantized weights always use the
826
+ # W8A8 kernel, due to lack of w8a16 support.
827
+ is_fp8_w8a8_supported = True
828
+
829
else:
830
is_fp8_w8a8_supported = self._check_scheme_supported(
831
CompressedTensorsW8A8Fp8.get_min_capability(), error=False