[ROCm][Bugfix] Convert ModelOpt FP8 per-channel weights to e4m3fnuz on MI300/MI325 (#47201)
Signed-off-by: Micah Williamson <micah.williamson@amd.com> Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
co-authored by
Andreas Karatzas
parent
6db31c8e76
commit
dd5c299fbe
@@ -249,10 +249,11 @@ def test_modelopt_fp8_pc_pt_checkpoint_setup(default_vllm_config, vllm_runner):
|
||||
assert isinstance(gate_up_proj.quant_method, ModelOptFp8PcPtLinearMethod)
|
||||
assert isinstance(down_proj.quant_method, ModelOptFp8PcPtLinearMethod)
|
||||
|
||||
assert qkv_proj.weight.dtype == torch.float8_e4m3fn
|
||||
assert o_proj.weight.dtype == torch.float8_e4m3fn
|
||||
assert gate_up_proj.weight.dtype == torch.float8_e4m3fn
|
||||
assert down_proj.weight.dtype == torch.float8_e4m3fn
|
||||
fp8_dtype = current_platform.fp8_dtype()
|
||||
assert qkv_proj.weight.dtype == fp8_dtype
|
||||
assert o_proj.weight.dtype == fp8_dtype
|
||||
assert gate_up_proj.weight.dtype == fp8_dtype
|
||||
assert down_proj.weight.dtype == fp8_dtype
|
||||
|
||||
# Per-channel scales; activations are dynamically scaled per token.
|
||||
assert hasattr(qkv_proj, "weight_scale")
|
||||
|
||||
Reference in New Issue
Block a user