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 14, 05:30 UTC
[BugFix] [Spec Decode] Remove LlamaForCausalLMEagle3 to fix CI (#22611)
Signed-off-by: 22quinn <33176974+22quinn@users.noreply.github.com>
22quinn committed
Aug 11, 2025 at 10:31 UTC
807d21b80d11437f10dc3360ad8215f3ca6eb2e8
4 files changed
+32
-27
tests/models/registry.py
+5
-4
@@ -525,10 +525,11 @@ _SPECULATIVE_DECODING_EXAMPLE_MODELS = {
525
trust_remote_code=True,
526
speculative_model="yuhuili/EAGLE3-LLaMA3.1-Instruct-8B",
527
tokenizer="meta-llama/Llama-3.1-8B-Instruct"),
528
- "LlamaForCausalLMEagle3": _HfExamplesInfo("AngelSlim/Qwen3-8B_eagle3", # noqa: E501
529
- trust_remote_code=True,
530
- speculative_model="AngelSlim/Qwen3-8B_eagle3",
531
- tokenizer="Qwen/Qwen3-8B"),
528
+ # TODO: Re-enable this once tests/models/test_initialization.py is fixed, see PR #22333 #22611 # noqa: E501
529
+ # "LlamaForCausalLMEagle3": _HfExamplesInfo("AngelSlim/Qwen3-8B_eagle3", # noqa: E501
530
+ # trust_remote_code=True,
531
+ # speculative_model="AngelSlim/Qwen3-8B_eagle3", # noqa: E501
532
+ # tokenizer="Qwen/Qwen3-8B"),
533
"EagleLlama4ForCausalLM": _HfExamplesInfo(
534
"morgendave/EAGLE-Llama-4-Scout-17B-16E-Instruct",
535
trust_remote_code=True,
tests/v1/e2e/test_spec_decode.py
+24
-21
@@ -125,27 +125,30 @@ def test_ngram_correctness(
125
cleanup_dist_env_and_memory()
126
127
128
-@pytest.mark.parametrize(["model_setup", "mm_enabled"], [
129
- (("eagle3", "Qwen/Qwen3-8B", "AngelSlim/Qwen3-8B_eagle3", 1), False),
130
- (("eagle", "meta-llama/Llama-3.1-8B-Instruct",
131
- "yuhuili/EAGLE-LLaMA3.1-Instruct-8B", 1), False),
132
- (("eagle3", "meta-llama/Llama-3.1-8B-Instruct",
133
- "yuhuili/EAGLE3-LLaMA3.1-Instruct-8B", 1), False),
134
- pytest.param(
135
- ("eagle", "meta-llama/Llama-4-Scout-17B-16E-Instruct",
136
- "morgendave/EAGLE-Llama-4-Scout-17B-16E-Instruct", 4),
137
- False,
138
- marks=pytest.mark.skip(reason="Skipping due to CI OOM issues")),
139
- pytest.param(
140
- ("eagle", "meta-llama/Llama-4-Scout-17B-16E-Instruct",
141
- "morgendave/EAGLE-Llama-4-Scout-17B-16E-Instruct", 4),
142
- True,
143
- marks=pytest.mark.skip(reason="Skipping due to CI OOM issues")),
144
-],
145
- ids=[
146
- "qwen3_eagle3", "llama3_eagle", "llama3_eagle3",
147
- "llama4_eagle", "llama4_eagle_mm"
148
- ])
128
+@pytest.mark.parametrize(
129
+ ["model_setup", "mm_enabled"],
130
+ [
131
+ # TODO: Re-enable this once tests/models/test_initialization.py is fixed, see PR #22333 #22611 # noqa: E501
132
+ # (("eagle3", "Qwen/Qwen3-8B", "AngelSlim/Qwen3-8B_eagle3", 1), False),
133
+ (("eagle", "meta-llama/Llama-3.1-8B-Instruct",
134
+ "yuhuili/EAGLE-LLaMA3.1-Instruct-8B", 1), False),
135
+ (("eagle3", "meta-llama/Llama-3.1-8B-Instruct",
136
+ "yuhuili/EAGLE3-LLaMA3.1-Instruct-8B", 1), False),
137
+ pytest.param(
138
+ ("eagle", "meta-llama/Llama-4-Scout-17B-16E-Instruct",
139
+ "morgendave/EAGLE-Llama-4-Scout-17B-16E-Instruct", 4),
140
+ False,
141
+ marks=pytest.mark.skip(reason="Skipping due to CI OOM issues")),
142
+ pytest.param(
143
+ ("eagle", "meta-llama/Llama-4-Scout-17B-16E-Instruct",
144
+ "morgendave/EAGLE-Llama-4-Scout-17B-16E-Instruct", 4),
145
+ True,
146
+ marks=pytest.mark.skip(reason="Skipping due to CI OOM issues")),
147
+ ],
148
+ ids=[
149
+ "qwen3_eagle3", "llama3_eagle", "llama3_eagle3", "llama4_eagle",
150
+ "llama4_eagle_mm"
151
+ ])
152
@pytest.mark.parametrize("attn_backend",
153
get_attn_backend_list_based_on_platform())
154
def test_eagle_correctness(
vllm/model_executor/models/registry.py
+2
-1
@@ -259,7 +259,8 @@ _SPECULATIVE_DECODING_MODELS = {
259
"EagleLlama4ForCausalLM": ("llama4_eagle", "EagleLlama4ForCausalLM"),
260
"EagleMiniCPMForCausalLM": ("minicpm_eagle", "EagleMiniCPMForCausalLM"),
261
"Eagle3LlamaForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
262
- "LlamaForCausalLMEagle3": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
262
+ # TODO: Re-enable this once tests/models/test_initialization.py is fixed, see PR #22333 #22611 # noqa: E501
263
+ # "LlamaForCausalLMEagle3": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
264
"DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
265
"Glm4MoeMTPModel": ("glm4_moe_mtp", "Glm4MoeMTP"),
266
"MedusaModel": ("medusa", "Medusa"),
vllm/transformers_utils/configs/eagle.py
+1
-1
@@ -45,7 +45,7 @@ class EAGLEConfig(PretrainedConfig):
45
46
# Eagle model name should follow naming convention of
47
# LlamaForCausalLM -> EagleLlamaForCausalLM
48
- # LlamaForCausalLM -> Eagle3LlamaForCausalLM / LlamaForCausalLMEagle3
48
+ # LlamaForCausalLM -> Eagle3LlamaForCausalLM
49
if method == "eagle":
50
assert self.model is not None, \
51
"model should not be None when method is eagle"