Merge remote-tracking branch 'upstream/main' into patches
This commit is contained in:
@@ -26,6 +26,7 @@ from vllm.config import SpeculativeConfig, VllmConfig
|
||||
from vllm.config.compilation import CUDAGraphMode
|
||||
from vllm.config.quantization import QuantizationConfigArgs
|
||||
from vllm.model_executor.layers.quantization.fp8 import Fp8Config
|
||||
from vllm.model_executor.layers.quantization.inc import INCConfig
|
||||
from vllm.model_executor.layers.quantization.modelopt import (
|
||||
ModelOptMixedPrecisionConfig,
|
||||
ModelOptNvFp4Config,
|
||||
@@ -573,6 +574,24 @@ def test_ple_embedding_respects_modelopt_exclusion() -> None:
|
||||
)
|
||||
|
||||
|
||||
def test_ple_embedding_respects_inc_layer_config() -> None:
|
||||
prefix = "model.layers.1.ple.ple_embedding.ngram_embedding"
|
||||
quant_config = INCConfig(
|
||||
weight_bits=4,
|
||||
group_size=128,
|
||||
block_name_to_quantize="model.layers",
|
||||
extra_config={".*ple.*": {"bits": 16, "data_type": "float"}},
|
||||
)
|
||||
assert isinstance(
|
||||
Qwen4ExpPLEEmbeddingMethod.from_quant_config(quant_config, prefix),
|
||||
Qwen4ExpPLEUnquantizedEmbeddingMethod,
|
||||
)
|
||||
|
||||
quant_config.extra_config = None
|
||||
with pytest.raises(NotImplementedError, match="INCConfig"):
|
||||
Qwen4ExpPLEEmbeddingMethod.from_quant_config(quant_config, prefix)
|
||||
|
||||
|
||||
def test_ple_embedding_dtype_overrides_modelopt_exclusion() -> None:
|
||||
prefix = "model.layers.1.ple.ple_embedding.ngram_embedding"
|
||||
quant_config = ModelOptNvFp4Config(exclude_modules=[prefix])
|
||||
|
||||
@@ -380,12 +380,9 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
|
||||
"Llama4ForCausalLM": _HfExamplesInfo(
|
||||
"meta-llama/Llama-4-Scout-17B-16E-Instruct",
|
||||
),
|
||||
"LongcatFlashForCausalLM": _HfExamplesInfo(
|
||||
"meituan-longcat/LongCat-Flash-Chat", trust_remote_code=True
|
||||
),
|
||||
"LongcatFlashForCausalLM": _HfExamplesInfo("meituan-longcat/LongCat-Flash-Chat"),
|
||||
"LongcatFlashNgramForCausalLM": _HfExamplesInfo(
|
||||
"meituan-longcat/LongCat-Flash-Lite",
|
||||
trust_remote_code=True,
|
||||
# Shrink the ~62GB n-gram tables (ngram_vocab_size_ratio * vocab_size)
|
||||
# so the dummy-weight init test fits in CI memory.
|
||||
hf_overrides={"ngram_vocab_size_ratio": 1},
|
||||
@@ -1752,7 +1749,6 @@ _SPECULATIVE_DECODING_EXAMPLE_MODELS = {
|
||||
),
|
||||
"LongCatFlashMTPModel": _HfExamplesInfo(
|
||||
"meituan-longcat/LongCat-Flash-Chat",
|
||||
trust_remote_code=True,
|
||||
speculative_model="meituan-longcat/LongCat-Flash-Chat",
|
||||
),
|
||||
"MiMoMTPModel": _HfExamplesInfo(
|
||||
|
||||
Reference in New Issue
Block a user