[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:
fxmarty-amd
2026-09-24 12:48:14 -07:00
committed by GitHub
co-authored by mgoin mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 58aa2f1ef1
commit 90a9515006
15 changed files with 240 additions and 269 deletions
@@ -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",
)
@@ -1158,13 +1158,13 @@ def test_kv_scale_reload(vllm_runner):
"Qwen/Qwen3-0.6B",
"inference-optimization/Qwen3-0.6B-debug-multiply",
"inference-optimization/Qwen3-0.6B-debug-add",
"fp8",
"fp8_per_tensor",
),
pytest.param(
"inference-optimization/DeepSeek-V3-debug-empty",
"inference-optimization/DeepSeek-V3-debug-multiply",
"inference-optimization/DeepSeek-V3-debug-add",
"fp8",
"fp8_per_tensor",
marks=[pytest.mark.slow_test],
),
pytest.param(
@@ -1192,7 +1192,7 @@ def test_online_quantize_reload(
if current_platform.device_count() < tp_size:
pytest.skip(reason="Not enough GPU devices")
if quantization == "fp8" and _fp8_reload_unsupported():
if quantization == "fp8_per_tensor" and _fp8_reload_unsupported():
pytest.skip(reason="Requires FP8 support")
with vllm_runner(