@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / e603129505

Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 12, 13:30 UTC

[refactor] CTConfig methods to static/class methods (#28870)

Signed-off-by: HDCharles <charlesdavidhernandez@gmail.com> Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>

HDCharles committed Nov 26, 2025 at 12:21 UTC e603129505fdf39b0784fe9600feb9101ed5170d

1 file changed +32 -23

vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py

+32 -23

@@ -266,8 +266,9 @@ class CompressedTensorsConfig(QuantizationConfig):

266 def get_config_filenames(cls) -> list[str]:

267 return []

268

269 + @staticmethod

270 def _check_scheme_supported(

270 - self, min_capability: int, error: bool = True, match_exact: bool = False

271 + min_capability: int, error: bool = True, match_exact: bool = False

272 ) -> bool:

273 capability_tuple = current_platform.get_device_capability()

274

@@ -293,9 +294,8 @@ class CompressedTensorsConfig(QuantizationConfig):

294 else:

295 return False

296

296 - def _is_fp4a4_nvfp4(

297 - self, weight_quant: QuantizationArgs, input_quant: QuantizationArgs

298 - ):

297 + @staticmethod

298 + def _is_fp4a4_nvfp4(weight_quant: QuantizationArgs, input_quant: QuantizationArgs):

299 if weight_quant is None or input_quant is None:

300 return False

301

@@ -322,9 +322,8 @@ class CompressedTensorsConfig(QuantizationConfig):

322 and is_symmetric

323 )

324

325 - def _is_fp4a16_nvfp4(

326 - self, weight_quant: QuantizationArgs, input_quant: QuantizationArgs

327 - ):

325 + @staticmethod

326 + def _is_fp4a16_nvfp4(weight_quant: QuantizationArgs, input_quant: QuantizationArgs):

327 is_weight_only = weight_quant is not None and input_quant is None

328 is_tensor_group_quant = (

329 weight_quant.strategy == QuantizationStrategy.TENSOR_GROUP.value

@@ -344,8 +343,9 @@ class CompressedTensorsConfig(QuantizationConfig):

343 and is_symmetric

344 )

345

346 + @staticmethod

347 def _is_static_tensor_w8a8(

348 - self, weight_quant: QuantizationArgs, input_quant: QuantizationArgs

348 + weight_quant: QuantizationArgs, input_quant: QuantizationArgs

349 ) -> bool:

350 is_8_bits = weight_quant.num_bits == input_quant.num_bits == 8

351 weight_strategy = (

@@ -362,8 +362,9 @@ class CompressedTensorsConfig(QuantizationConfig):

362 # Only symmetric weight quantization supported.

363 return is_8_bits and is_tensor and weight_quant.symmetric and is_static

364

365 + @staticmethod

366 def _is_dynamic_token_w8a8(

366 - self, weight_quant: QuantizationArgs, input_quant: QuantizationArgs

367 + weight_quant: QuantizationArgs, input_quant: QuantizationArgs

368 ) -> bool:

369 is_8_bits = weight_quant.num_bits == input_quant.num_bits == 8

370 weight_strategy = (

@@ -379,8 +380,9 @@ class CompressedTensorsConfig(QuantizationConfig):

380 # Only symmetric weight quantization supported.

381 return is_8_bits and is_token and weight_quant.symmetric and is_dynamic

382

383 + @staticmethod

384 def _is_dynamic_token_w4a8_int(

383 - self, weight_quant: QuantizationArgs, input_quant: QuantizationArgs

385 + weight_quant: QuantizationArgs, input_quant: QuantizationArgs

386 ) -> bool:

387 is_weight_4_bits = weight_quant.num_bits == 4

388 is_activation_8_bits = input_quant.num_bits == 8

@@ -403,8 +405,9 @@ class CompressedTensorsConfig(QuantizationConfig):

405 and is_dynamic

406 )

407

408 + @staticmethod

409 def _is_fp8_w8a8(

407 - self, weight_quant: QuantizationArgs, input_quant: QuantizationArgs

410 + weight_quant: QuantizationArgs, input_quant: QuantizationArgs

411 ) -> bool:

412 # Confirm weights and activations quantized.

413 if weight_quant is None or input_quant is None:

@@ -439,8 +442,9 @@ class CompressedTensorsConfig(QuantizationConfig):

442 is_per_tensor_activation = input_quant.strategy == QuantizationStrategy.TENSOR

443 return is_symmetric_activation and is_per_tensor_activation

444

445 + @staticmethod

446 def _is_fp8_w4a8(

443 - self, weight_quant: QuantizationArgs, input_quant: QuantizationArgs

447 + weight_quant: QuantizationArgs, input_quant: QuantizationArgs

448 ) -> bool:

449 if not weight_quant or not input_quant:

450 return False

@@ -462,29 +466,33 @@ class CompressedTensorsConfig(QuantizationConfig):

466 and is_dynamic

467 )

468

469 + @classmethod

470 def _is_fp8_w4a8_sm90(

466 - self, weight_quant: QuantizationArgs, input_quant: QuantizationArgs

471 + cls, weight_quant: QuantizationArgs, input_quant: QuantizationArgs

472 ) -> bool:

468 - return self._check_scheme_supported(

473 + return cls._check_scheme_supported(

474 90, error=False, match_exact=True

470 - ) and self._is_fp8_w4a8(weight_quant, input_quant)

475 + ) and cls._is_fp8_w4a8(weight_quant, input_quant)

476

477 + @classmethod

478 def _is_fp8_w8a8_sm90(

473 - self, weight_quant: QuantizationArgs, input_quant: QuantizationArgs

479 + cls, weight_quant: QuantizationArgs, input_quant: QuantizationArgs

480 ) -> bool:

475 - return self._check_scheme_supported(

481 + return cls._check_scheme_supported(

482 90, error=False, match_exact=True

477 - ) and self._is_fp8_w8a8(weight_quant, input_quant)

483 + ) and cls._is_fp8_w8a8(weight_quant, input_quant)

484

485 + @classmethod

486 def _is_fp8_w8a8_sm100(

480 - self, weight_quant: QuantizationArgs, input_quant: QuantizationArgs

487 + cls, weight_quant: QuantizationArgs, input_quant: QuantizationArgs

488 ) -> bool:

482 - return self._check_scheme_supported(

489 + return cls._check_scheme_supported(

490 100, error=False, match_exact=True

484 - ) and self._is_fp8_w8a8(weight_quant, input_quant)

491 + ) and cls._is_fp8_w8a8(weight_quant, input_quant)

492

493 + @staticmethod

494 def _is_fp8_w8a16(

487 - self, weight_quant: QuantizationArgs, input_quant: QuantizationArgs

495 + weight_quant: QuantizationArgs, input_quant: QuantizationArgs

496 ) -> bool:

497 # Confirm weights quantized.

498 if weight_quant is None:

@@ -508,8 +516,9 @@ class CompressedTensorsConfig(QuantizationConfig):

516 and is_tensor_or_channel_or_block_weight

517 )

518

519 + @staticmethod

520 def _is_wNa16_group_channel(

512 - self, weight_quant: QuantizationArgs, input_quant: QuantizationArgs

521 + weight_quant: QuantizationArgs, input_quant: QuantizationArgs

522 ) -> bool:

523 input_quant_none = input_quant is None

524 is_channel_group = (