[Cleanup] Remove online quantization support in fp8.py in favor of online shorthands (#53585)
Signed-off-by: Felix Marty <Felix.Marty@amd.com> Signed-off-by: mgoin <mgoin64@gmail.com> Co-authored-by: mgoin <mgoin64@gmail.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
co-authored by
mgoin
mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent
58aa2f1ef1
commit
90a9515006
@@ -1227,7 +1227,6 @@ def test_non_quark_shared_expert_fse_is_incompatible() -> None:
|
||||
|
||||
def _fp8_config(**kwargs: Any) -> Fp8Config:
|
||||
return Fp8Config(
|
||||
is_checkpoint_fp8_serialized=True,
|
||||
activation_scheme="dynamic",
|
||||
weight_block_size=[128, 128],
|
||||
**kwargs,
|
||||
@@ -1248,7 +1247,7 @@ def test_block_fp8_shared_expert_fse_is_compatible() -> None:
|
||||
def test_per_tensor_fp8_shared_expert_fse_is_incompatible() -> None:
|
||||
"""Per-tensor scales are 0-D, so the shared-expert chunker cannot slice them."""
|
||||
compatible, reason = is_shared_expert_quant_fse_compatible(
|
||||
Fp8Config(is_checkpoint_fp8_serialized=True, activation_scheme="dynamic"),
|
||||
Fp8Config(activation_scheme="dynamic"),
|
||||
"model.layers.0.mlp.experts",
|
||||
"model.layers.0.mlp.shared_experts",
|
||||
)
|
||||
|
||||
Reference in New Issue
Block a user