[ROCm][Test] Use platform FP8 dtype in ModelOpt FP8_PB_WO test (#53268)
Signed-off-by: Djordje Ramic <djoramic@amd.com>
djramic committed
Aug 21, 2026 at 22:06 UTC
f15ea66b3006ac3b6ea1571248c07f7b35916dd2
1 file changed
+5
-4
tests/quantization/test_modelopt.py
+5
-4
@@ -508,10 +508,11 @@ def test_modelopt_fp8_pb_wo_checkpoint_setup(default_vllm_config, vllm_runner):
508
assert isinstance(gate_up_proj.quant_method, ModelOptFp8PbWoLinearMethod)
509
assert isinstance(down_proj.quant_method, ModelOptFp8PbWoLinearMethod)
510
511
- assert qkv_proj.weight.dtype == torch.float8_e4m3fn
512
- assert o_proj.weight.dtype == torch.float8_e4m3fn
513
- assert gate_up_proj.weight.dtype == torch.float8_e4m3fn
514
- assert down_proj.weight.dtype == torch.float8_e4m3fn
511
+ fp8_dtype = current_platform.fp8_dtype()
512
+ assert qkv_proj.weight.dtype == fp8_dtype
513
+ assert o_proj.weight.dtype == fp8_dtype
514
+ assert gate_up_proj.weight.dtype == fp8_dtype
515
+ assert down_proj.weight.dtype == fp8_dtype
516
517
# Block scales; should be materialized as a 2D [out_blk, in_blk] tensor.
518
assert hasattr(qkv_proj, "weight_scale")