xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-09 18:25:48 +08:00
Code Issues Packages Projects Releases Wiki Activity

[BugFix] [Spec Decode] Remove LlamaForCausalLMEagle3 to fix CI (#22611)

Browse Source 
Signed-off-by: 22quinn <33176974+22quinn@users.noreply.github.com>
This commit is contained in:
22quinn 2025-08-11 10:31:36 -07:00 committed by GitHub
parent c90fb03df5
commit 807d21b80d
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
4 changed files with 32 additions and 27 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
tests/models/registry.py
View File

@ -525,10 +525,11 @@ _SPECULATIVE_DECODING_EXAMPLE_MODELS = {
trust_remote_code=True,
speculative_model="yuhuili/EAGLE3-LLaMA3.1-Instruct-8B",
tokenizer="meta-llama/Llama-3.1-8B-Instruct"),
"LlamaForCausalLMEagle3": _HfExamplesInfo("AngelSlim/Qwen3-8B_eagle3", # noqa: E501
trust_remote_code=True,
speculative_model="AngelSlim/Qwen3-8B_eagle3",
tokenizer="Qwen/Qwen3-8B"),
# TODO: Re-enable this once tests/models/test_initialization.py is fixed, see PR #22333 #22611 # noqa: E501
# "LlamaForCausalLMEagle3": _HfExamplesInfo("AngelSlim/Qwen3-8B_eagle3", # noqa: E501
# trust_remote_code=True,
# speculative_model="AngelSlim/Qwen3-8B_eagle3", # noqa: E501
# tokenizer="Qwen/Qwen3-8B"),
"EagleLlama4ForCausalLM": _HfExamplesInfo(
"morgendave/EAGLE-Llama-4-Scout-17B-16E-Instruct",
trust_remote_code=True,

45
tests/v1/e2e/test_spec_decode.py
View File

@ -125,27 +125,30 @@ def test_ngram_correctness(
cleanup_dist_env_and_memory()
@pytest.mark.parametrize(["model_setup", "mm_enabled"], [
(("eagle3", "Qwen/Qwen3-8B", "AngelSlim/Qwen3-8B_eagle3", 1), False),
(("eagle", "meta-llama/Llama-3.1-8B-Instruct",
"yuhuili/EAGLE-LLaMA3.1-Instruct-8B", 1), False),
(("eagle3", "meta-llama/Llama-3.1-8B-Instruct",
"yuhuili/EAGLE3-LLaMA3.1-Instruct-8B", 1), False),
pytest.param(
("eagle", "meta-llama/Llama-4-Scout-17B-16E-Instruct",
"morgendave/EAGLE-Llama-4-Scout-17B-16E-Instruct", 4),
False,
marks=pytest.mark.skip(reason="Skipping due to CI OOM issues")),
pytest.param(
("eagle", "meta-llama/Llama-4-Scout-17B-16E-Instruct",
"morgendave/EAGLE-Llama-4-Scout-17B-16E-Instruct", 4),
True,
marks=pytest.mark.skip(reason="Skipping due to CI OOM issues")),
],
ids=[
"qwen3_eagle3", "llama3_eagle", "llama3_eagle3",
"llama4_eagle", "llama4_eagle_mm"
])
@pytest.mark.parametrize(
["model_setup", "mm_enabled"],
[
# TODO: Re-enable this once tests/models/test_initialization.py is fixed, see PR #22333 #22611 # noqa: E501
# (("eagle3", "Qwen/Qwen3-8B", "AngelSlim/Qwen3-8B_eagle3", 1), False),
(("eagle", "meta-llama/Llama-3.1-8B-Instruct",
"yuhuili/EAGLE-LLaMA3.1-Instruct-8B", 1), False),
(("eagle3", "meta-llama/Llama-3.1-8B-Instruct",
"yuhuili/EAGLE3-LLaMA3.1-Instruct-8B", 1), False),
pytest.param(
("eagle", "meta-llama/Llama-4-Scout-17B-16E-Instruct",
"morgendave/EAGLE-Llama-4-Scout-17B-16E-Instruct", 4),
False,
marks=pytest.mark.skip(reason="Skipping due to CI OOM issues")),
pytest.param(
("eagle", "meta-llama/Llama-4-Scout-17B-16E-Instruct",
"morgendave/EAGLE-Llama-4-Scout-17B-16E-Instruct", 4),
True,
marks=pytest.mark.skip(reason="Skipping due to CI OOM issues")),
],
ids=[
"qwen3_eagle3", "llama3_eagle", "llama3_eagle3", "llama4_eagle",
"llama4_eagle_mm"
])
@pytest.mark.parametrize("attn_backend",
get_attn_backend_list_based_on_platform())
def test_eagle_correctness(

3
vllm/model_executor/models/registry.py
View File

@ -259,7 +259,8 @@ _SPECULATIVE_DECODING_MODELS = {
"EagleLlama4ForCausalLM": ("llama4_eagle", "EagleLlama4ForCausalLM"),
"EagleMiniCPMForCausalLM": ("minicpm_eagle", "EagleMiniCPMForCausalLM"),
"Eagle3LlamaForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
"LlamaForCausalLMEagle3": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
# TODO: Re-enable this once tests/models/test_initialization.py is fixed, see PR #22333 #22611 # noqa: E501
# "LlamaForCausalLMEagle3": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
"DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
"Glm4MoeMTPModel": ("glm4_moe_mtp", "Glm4MoeMTP"),
"MedusaModel": ("medusa", "Medusa"),

2
vllm/transformers_utils/configs/eagle.py
View File

@ -45,7 +45,7 @@ class EAGLEConfig(PretrainedConfig):
# Eagle model name should follow naming convention of
# LlamaForCausalLM -> EagleLlamaForCausalLM
# LlamaForCausalLM -> Eagle3LlamaForCausalLM / LlamaForCausalLMEagle3
# LlamaForCausalLM -> Eagle3LlamaForCausalLM
if method == "eagle":
assert self.model is not None, \
"model should not be None when method is eagle"