Upstream Llama4 Support to Main (#16113)

Signed-off-by: Aston Zhang <[email protected]>
Signed-off-by: Chris Thi <[email protected]>
Signed-off-by: drisspg <[email protected]>
Signed-off-by: Jon Swenson <[email protected]>
Signed-off-by: Keyun Tong <[email protected]>
Signed-off-by: Lu Fang <[email protected]>
Signed-off-by: Xiaodong Wang <[email protected]>
Signed-off-by: Yang Chen <[email protected]>
Signed-off-by: Ye (Charlotte) Qi <[email protected]>
Signed-off-by: Yong Hoon Shin <[email protected]>
Signed-off-by: Zijing Liu <[email protected]>
Signed-off-by: Lu Fang <[email protected]>
Signed-off-by: Lu Fang <[email protected]>
Signed-off-by: Lucia Fang <[email protected]>
Signed-off-by: Roger Wang <[email protected]>
Signed-off-by: DarkLight1337 <[email protected]>
Co-authored-by: Lu Fang <[email protected]>
Co-authored-by: Roger Wang <[email protected]>
Co-authored-by: DarkLight1337 <[email protected]>

Author: 4 people

.buildkite/test-pipeline.yaml

@@ -389,7 +389,8 @@ steps:
     - pytest -v -s models/test_transformers.py
     - pytest -v -s models/test_registry.py
     # V1 Test: https://github.com/vllm-project/vllm/issues/14531
-    - VLLM_USE_V1=0 pytest -v -s models/test_initialization.py
+    - VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'not llama4'
+    - VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'llama4'
 
 - label: Language Models Test (Standard) # 32min
   #mirror_hardwares: [amd]

benchmarks/kernels/benchmark_moe.py

@@ -553,6 +553,9 @@ def main(args: argparse.Namespace):
         intermediate_size = config.moe_intermediate_size
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
     else:
+        if not hasattr(config, "hidden_size"):
+            # Support for llama4
+            config = config.text_config
         # Default: Mixtral.
         E = config.num_local_experts
         topk = config.num_experts_per_tok

docs/source/models/supported_models.md

@@ -24,7 +24,7 @@ vLLM also supports model implementations that are available in Transformers. Thi
 
 To check if the modeling backend is Transformers, you can simply do this:
 
-```python 
+```python
 from vllm import LLM
 llm = LLM(model=..., task="generate")  # Name or path of your model
 llm.apply_model(lambda model: print(type(model)))
@@ -55,7 +55,7 @@ If your model is neither supported natively by vLLM or Transformers, you can sti
 Simply set `trust_remote_code=True` and vLLM will run any model on the Model Hub that is compatible with Transformers.
 Provided that the model writer implements their model in a compatible way, this means that you can run new models before they are officially supported in Transformers or vLLM!
 
-```python 
+```python
 from vllm import LLM
 llm = LLM(model=..., task="generate", trust_remote_code=True)  # Name or path of your model
 llm.apply_model(lambda model: print(model.__class__))
@@ -850,6 +850,13 @@ See [this page](#generative-models) for more information on how to use generativ
   *
   * ✅︎
   * ✅︎
+- * `Llama4ForConditionalGeneration`
+  * Llama-4-17B-Omni-Instruct
+  * T + I<sup>+</sup>
+  * `meta-llama/Llama-4-Scout-17B-16E-Instruct`, `meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8`, `meta-llama/Llama-4-Maverick-17B-128E-Instruct`, etc.
+  *
+  * ✅︎
+  * ✅︎
 - * `LlavaForConditionalGeneration`
   * LLaVA-1.5
   * T + I<sup>E+</sup>

examples/offline_inference/audio_language.py

@@ -47,7 +47,7 @@ def run_minicpmo(question: str, audio_count: int) -> ModelRequestData:
         model=model_name,
         trust_remote_code=True,
         max_model_len=4096,
-        max_num_seqs=5,
+        max_num_seqs=2,
         limit_mm_per_prompt={"audio": audio_count},
     )
 

examples/offline_inference/vision_language.py

@@ -582,6 +582,42 @@ def run_mllama(questions: list[str], modality: str) -> ModelRequestData:
     )
 
 
+def run_llama4(questions: list[str], modality: str):
+    assert modality == "image"
+
+    model_name = "meta-llama/Llama-4-Scout-17B-16E-Instruct"
+
+    engine_args = EngineArgs(
+        model=model_name,
+        max_model_len=8192,
+        max_num_seqs=4,
+        tensor_parallel_size=8,
+        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        gpu_memory_utilization=0.4,
+    )
+
+    tokenizer = AutoTokenizer.from_pretrained(model_name)
+    messages = [[{
+        "role":
+        "user",
+        "content": [{
+            "type": "image"
+        }, {
+            "type": "text",
+            "text": f"{question}"
+        }]
+    }] for question in questions]
+    prompts = tokenizer.apply_chat_template(messages,
+                                            add_generation_prompt=True,
+                                            tokenize=False)
+    stop_token_ids = None
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompts=prompts,
+        stop_token_ids=stop_token_ids,
+    )
+
+
 # Molmo
 def run_molmo(questions: list[str], modality: str) -> ModelRequestData:
     assert modality == "image"
@@ -907,6 +943,7 @@ def run_skyworkr1v(questions: list[str], modality: str) -> ModelRequestData:
     "minicpmv": run_minicpmv,
     "mistral3": run_mistral3,
     "mllama": run_mllama,
+    "llama4": run_llama4,
     "molmo": run_molmo,
     "NVLM_D": run_nvlm_d,
     "paligemma": run_paligemma,

examples/offline_inference/vision_language_multi_image.py

@@ -253,6 +253,43 @@ def load_internvl(question: str, image_urls: list[str]) -> ModelRequestData:
     )
 
 
+def load_llama4(question: str, image_urls: list[str]) -> ModelRequestData:
+    model_name = "meta-llama/Llama-4-Scout-17B-16E-Instruct"
+
+    engine_args = EngineArgs(
+        model=model_name,
+        max_model_len=8192,
+        max_num_seqs=4,
+        tensor_parallel_size=8,
+        limit_mm_per_prompt={"image": len(image_urls)},
+    )
+
+    placeholders = [{"type": "image", "image": url} for url in image_urls]
+    messages = [{
+        "role":
+        "user",
+        "content": [
+            *placeholders,
+            {
+                "type": "text",
+                "text": question
+            },
+        ],
+    }]
+
+    processor = AutoProcessor.from_pretrained(model_name)
+
+    prompt = processor.apply_chat_template(messages,
+                                           tokenize=False,
+                                           add_generation_prompt=True)
+
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompt=prompt,
+        image_data=[fetch_image(url) for url in image_urls],
+    )
+
+
 def load_mistral3(question: str, image_urls: list[str]) -> ModelRequestData:
     model_name = "mistralai/Mistral-Small-3.1-24B-Instruct-2503"
 
@@ -567,6 +604,7 @@ def load_qwen2_5_vl(question: str, image_urls: list[str]) -> ModelRequestData:
     "h2ovl_chat": load_h2ovl,
     "idefics3": load_idefics3,
     "internvl_chat": load_internvl,
+    "llama4": load_llama4,
     "mistral3": load_mistral3,
     "mllama": load_mllama,
     "NVLM_D": load_nvlm_d,

requirements/common.txt

@@ -6,7 +6,7 @@ requests >= 2.26.0
 tqdm
 blake3
 py-cpuinfo
-transformers >= 4.50.3
+transformers >= 4.51.0
 huggingface-hub[hf_xet] >= 0.30.0  # Required for Xet downloads.
 tokenizers >= 0.19.1  # Required for Llama 3.
 protobuf # Required by LlamaTokenizer.

requirements/test.in

@@ -30,7 +30,7 @@ mistral_common[opencv] >= 1.5.4 # required for pixtral test
 opencv-python-headless >= 4.11.0 # required for video test
 datamodel_code_generator # required for minicpm3 test
 lm-eval[api]==0.4.8 # required for model evaluation test
-transformers==4.50.3
+transformers==4.51.0
 huggingface-hub[hf_xet]>=0.30.0  # Required for Xet downloads.
 # quantization
 bitsandbytes>=0.45.3

requirements/test.txt

@@ -645,7 +645,7 @@ tqdm==4.66.6
     #   transformers
 tqdm-multiprocess==0.0.11
     # via lm-eval
-transformers==4.50.3
+transformers==4.51.0
     # via
     #   -r requirements/test.in
     #   genai-perf

tests/models/decoder_only/audio_language/test_ultravox.py

@@ -12,6 +12,7 @@
 
 from ....conftest import HfRunner, VllmRunner
 from ....utils import RemoteOpenAIServer
+from ...registry import HF_EXAMPLE_MODELS
 from ...utils import check_logprobs_close
 
 MODEL_NAME = "fixie-ai/ultravox-v0_5-llama-3_2-1b"
@@ -55,7 +56,10 @@ def server(request, audio_assets):
         for key, value in request.param.items()
     ]
 
-    with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
+    with RemoteOpenAIServer(MODEL_NAME,
+                            args,
+                            env_dict={"VLLM_AUDIO_FETCH_TIMEOUT":
+                                      "30"}) as remote_server:
         yield remote_server
 
 
@@ -106,6 +110,10 @@ def run_test(
     **kwargs,
 ):
     """Inference result should be the same between hf and vllm."""
+    model_info = HF_EXAMPLE_MODELS.find_hf_info(model)
+    model_info.check_available_online(on_fail="skip")
+    model_info.check_transformers_version(on_fail="skip")
+
     # NOTE: take care of the order. run vLLM first, and then run HF.
     # vLLM needs a fresh new process without cuda initialization.
     # if we run HF first, the cuda initialization will be done and it
@@ -156,6 +164,10 @@ def run_multi_audio_test(
     num_logprobs: int,
     **kwargs,
 ):
+    model_info = HF_EXAMPLE_MODELS.find_hf_info(model)
+    model_info.check_available_online(on_fail="skip")
+    model_info.check_transformers_version(on_fail="skip")
+
     with vllm_runner(model,
                      dtype=dtype,
                      enforce_eager=True,