[Meta] Llama4 EAGLE Support (#20591)

Signed-off-by: qizixi <qizixi@meta.com>
Co-authored-by: qizixi <qizixi@meta.com>

examples/offline_inference/spec_decode.py

@@ -84,6 +84,7 @@ def main():
         gpu_memory_utilization=0.8,
         speculative_config=speculative_config,
         disable_log_stats=False,
+        max_model_len=16384,
     )
 
     sampling_params = SamplingParams(temperature=args.temp, max_tokens=args.output_len)

tests/models/registry.py

@@ -465,6 +465,11 @@ _SPECULATIVE_DECODING_EXAMPLE_MODELS = {
                                             trust_remote_code=True,
                                             speculative_model="yuhuili/EAGLE3-LLaMA3.1-Instruct-8B",
                                             tokenizer="meta-llama/Llama-3.1-8B-Instruct"),
+    "EagleLlama4ForCausalLM": _HfExamplesInfo(
+        "morgendave/EAGLE-Llama-4-Scout-17B-16E-Instruct",
+        trust_remote_code=True,
+        speculative_model="morgendave/EAGLE-Llama-4-Scout-17B-16E-Instruct",
+        tokenizer="meta-llama/Llama-4-Scout-17B-16E-Instruct"),  # noqa: E501
     "EagleMiniCPMForCausalLM": _HfExamplesInfo("openbmb/MiniCPM-1B-sft-bf16",
                                             trust_remote_code=True,
                                             is_available_online=False,

tests/models/test_initialization.py

@@ -36,6 +36,11 @@ def test_can_initialize(model_arch: str, monkeypatch: pytest.MonkeyPatch):
                       "KimiVLForConditionalGeneration"):
         pytest.skip("Avoid OOM")
 
+    if model_arch in ("Llama4ForCausalLM", "EagleLlama4ForCausalLM"):
+        from vllm.model_executor.models.llama4 import Llama4ForCausalLM
+        from vllm.model_executor.models.registry import ModelRegistry
+        ModelRegistry.register_model("Llama4ForCausalLM", Llama4ForCausalLM)
+
     # Avoid OOM and reduce initialization time by only using 1 layer
     def hf_overrides(hf_config: PretrainedConfig) -> PretrainedConfig:
         hf_config.update(model_info.hf_overrides)

tests/v1/e2e/test_spec_decode.py

@@ -6,8 +6,10 @@ import random
 from typing import Any
 
 import pytest
+import torch
 
 from vllm import LLM, SamplingParams
+from vllm.distributed import cleanup_dist_env_and_memory
 
 
 @pytest.fixture
@@ -53,14 +55,6 @@ def model_name():
     return "meta-llama/Llama-3.1-8B-Instruct"
 
 
-def eagle_model_name():
-    return "yuhuili/EAGLE-LLaMA3.1-Instruct-8B"
-
-
-def eagle3_model_name():
-    return "yuhuili/EAGLE3-LLaMA3.1-Instruct-8B"
-
-
 def test_ngram_correctness(
     monkeypatch: pytest.MonkeyPatch,
     test_prompts: list[list[dict[str, Any]]],
@@ -77,6 +71,8 @@ def test_ngram_correctness(
         ref_llm = LLM(model=model_name, max_model_len=1024)
         ref_outputs = ref_llm.chat(test_prompts, sampling_config)
         del ref_llm
+        torch.cuda.empty_cache()
+        cleanup_dist_env_and_memory()
 
         spec_llm = LLM(
             model=model_name,
@@ -103,34 +99,50 @@ def test_ngram_correctness(
         # Upon failure, inspect the outputs to check for inaccuracy.
         assert matches > int(0.7 * len(ref_outputs))
         del spec_llm
-
-
-@pytest.mark.parametrize("use_eagle3", [False, True], ids=["eagle", "eagle3"])
+        torch.cuda.empty_cache()
+        cleanup_dist_env_and_memory()
+
+
+@pytest.mark.parametrize("model_setup", [
+    ("eagle", "meta-llama/Llama-3.1-8B-Instruct",
+     "yuhuili/EAGLE-LLaMA3.1-Instruct-8B", 1),
+    ("eagle3", "meta-llama/Llama-3.1-8B-Instruct",
+     "yuhuili/EAGLE3-LLaMA3.1-Instruct-8B", 1),
+    pytest.param(
+        ("eagle", "meta-llama/Llama-4-Scout-17B-16E-Instruct",
+         "morgendave/EAGLE-Llama-4-Scout-17B-16E-Instruct", 4),
+        marks=pytest.mark.skip(reason="Skipping due to CI OOM issues")),
+],
+                         ids=["llama3_eagle", "llama3_eagle3", "llama4_eagle"])
 def test_eagle_correctness(
     monkeypatch: pytest.MonkeyPatch,
     test_prompts: list[list[dict[str, Any]]],
     sampling_config: SamplingParams,
-    model_name: str,
-    use_eagle3: bool,
+    model_setup: tuple[str, str, str, int],
 ):
     '''
     Compare the outputs of a original LLM and a speculative LLM
     should be the same when using eagle speculative decoding.
+    model_setup: (method, model_name, eagle_model_name, tp_size)
     '''
     with monkeypatch.context() as m:
         m.setenv("VLLM_USE_V1", "1")
+        method, model_name, spec_model_name, tp_size = model_setup
 
-        ref_llm = LLM(model=model_name, max_model_len=2048)
+        ref_llm = LLM(model=model_name,
+                      max_model_len=2048,
+                      tensor_parallel_size=tp_size)
         ref_outputs = ref_llm.chat(test_prompts, sampling_config)
         del ref_llm
+        torch.cuda.empty_cache()
+        cleanup_dist_env_and_memory()
 
-        spec_model_name = eagle3_model_name(
-        ) if use_eagle3 else eagle_model_name()
         spec_llm = LLM(
             model=model_name,
             trust_remote_code=True,
+            tensor_parallel_size=tp_size,
             speculative_config={
-                "method": "eagle3" if use_eagle3 else "eagle",
+                "method": method,
                 "model": spec_model_name,
                 "num_speculative_tokens": 3,
                 "max_model_len": 2048,
@@ -152,3 +164,5 @@ def test_eagle_correctness(
         # Upon failure, inspect the outputs to check for inaccuracy.
         assert matches > int(0.66 * len(ref_outputs))
         del spec_llm
+        torch.cuda.empty_cache()
+        cleanup_dist_env_and_memory()

vllm/model_executor/models/llama4_eagle.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# Copyright 2025 the LLAMA4, Meta Inc., vLLM, and HuggingFace Inc. team.
+# All rights reserved.
+#
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from collections.abc import Iterable
+from typing import Optional
+
+import torch
+import torch.nn as nn
+
+from vllm.compilation.decorators import support_torch_compile
+from vllm.config import VllmConfig
+from vllm.distributed.parallel_state import get_pp_group
+from vllm.logger import init_logger
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.logits_processor import LogitsProcessor
+from vllm.model_executor.layers.quantization.base_config import (
+    QuantizationConfig)
+from vllm.model_executor.layers.quantization.torchao import TorchAOConfig
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    VocabParallelEmbedding)
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+from vllm.model_executor.models.llama4 import (Llama4DecoderLayer,
+                                               Llama4ForCausalLM)
+from vllm.model_executor.models.utils import extract_layer_index
+
+from .utils import AutoWeightsLoader, maybe_prefix
+
+logger = init_logger(__name__)
+
+
+@support_torch_compile
+class LlamaModel(nn.Module):
+
+    def __init__(
+        self,
+        *,
+        vllm_config: VllmConfig,
+        prefix: str = "",
+        start_layer_id: int = 0,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.config = (
+            vllm_config.speculative_config.draft_model_config.hf_config)
+        self.validate_and_update_config(start_layer_id, quant_config)
+        self.vocab_size = self.config.vocab_size
+        self.embed_tokens = VocabParallelEmbedding(
+            self.config.vocab_size,
+            self.config.hidden_size,
+            prefix=maybe_prefix(prefix, "embed_tokens"),
+        )
+
+        self.layers = nn.ModuleList([
+            Llama4DecoderLayer(
+                self.config,
+                quant_config=quant_config,
+                prefix=maybe_prefix(prefix, f"layers.{i + start_layer_id}"),
+            ) for i in range(self.config.num_hidden_layers)
+        ])
+        self.fc = torch.nn.Linear(self.config.hidden_size * 2,
+                                  self.config.hidden_size,
+                                  bias=False)
+        self.norm = RMSNorm(self.config.hidden_size,
+                            eps=self.config.rms_norm_eps)
+
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor],
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        input_embeds = self.embed_tokens(input_ids)
+        hidden_states = self.fc(
+            torch.cat((input_embeds, hidden_states), dim=-1))
+        residual = None
+        for layer in self.layers:
+            hidden_states, residual = layer(
+                positions,
+                hidden_states,
+                residual,
+            )
+        hidden_states, _ = self.norm(hidden_states, residual)
+        return hidden_states, hidden_states
+
+    def load_weights(self, weights: Iterable[tuple[str,
+                                                   torch.Tensor]]) -> set[str]:
+        stacked_params_mapping = [
+            # (param_name, shard_name, shard_id)
+            (".qkv_proj", ".q_proj", "q"),
+            (".qkv_proj", ".k_proj", "k"),
+            (".qkv_proj", ".v_proj", "v"),
+            (".gate_up_proj", ".gate_proj", 0),
+            (".gate_up_proj", ".up_proj", 1),
+        ]
+        params_dict = dict(self.named_parameters())
+        loaded_params: set[str] = set()
+        for name, loaded_weight in weights:
+            name = name.removeprefix("model.")
+            for param_name, weight_name, shard_id in stacked_params_mapping:
+                if weight_name not in name:
+                    continue
+                name = name.replace(weight_name, param_name)
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                # if PP disabled then draft will share embed with target
+                if get_pp_group().world_size == 1 and \
+                    "embed_tokens." in name:
+                    continue
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader",
+                                        default_weight_loader)
+                weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        for name in params_dict:
+            # if PP disabled then draft will share embed with target
+            if get_pp_group().world_size == 1 and \
+                "embed_tokens." in name:
+                continue
+            assert name in loaded_params, f"{name} is not loaded!"
+        return loaded_params
+
+    def validate_and_update_config(
+            self,
+            start_layer_id: int,
+            quant_config: Optional[QuantizationConfig] = None) -> None:
+        # yoco and moe is not supported by draft model yet
+        assert self.config.yoco_global_kv_layer is None
+        assert self.config.yoco_local_kv_layer is None
+        assert len(self.config.moe_layers) == 0
+        # draft model layer index is increased by start_layer_id,
+        # so we need to pad relevant configs accordingly
+        self.config.no_rope_layers = [
+            0
+        ] * start_layer_id + self.config.no_rope_layers
+        # currently only TorchAO quantization is supported
+        if isinstance(quant_config, TorchAOConfig):
+
+            def pad_layer_name(layer: str) -> str:
+                layer_index = extract_layer_index(layer)
+                return layer.replace(str(layer_index),
+                                     str(layer_index + start_layer_id))
+
+            quant_config.torchao_config.module_fqn_to_config = {
+                pad_layer_name(layer): quantization
+                for layer, quantization in
+                quant_config.torchao_config.module_fqn_to_config.items()
+            }
+
+
+class EagleLlama4ForCausalLM(Llama4ForCausalLM):
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        nn.Module.__init__(self)
+        self.config = (
+            vllm_config.speculative_config.draft_model_config.hf_config)
+        target_layer_num = vllm_config.model_config.get_num_layers(
+            vllm_config.parallel_config)
+        # draft model quantization config may differ from target model
+        quant_config = VllmConfig.get_quantization_config(
+            vllm_config.speculative_config.draft_model_config,
+            vllm_config.load_config)
+        self.model = LlamaModel(vllm_config=vllm_config,
+                                prefix="model",
+                                start_layer_id=target_layer_num,
+                                quant_config=quant_config)
+        logit_scale = getattr(self.config, "logit_scale", 1.0)
+        self.logits_processor = LogitsProcessor(self.config.vocab_size,
+                                                scale=logit_scale)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        return self.model(input_ids, positions, hidden_states)
+
+    def load_weights(self, weights: Iterable[tuple[str,
+                                                   torch.Tensor]]) -> None:
+        loader = AutoWeightsLoader(
+            self,
+            # lm_head is tied with target model (Llama4ForCausalLM)
+            skip_prefixes=(["lm_head."]),
+        )
+
+        model_weights = {}
+        weights = [
+            self.permute_qk_weight_for_rotary(name, loaded_weight)
+            for name, loaded_weight in weights
+        ]
+        for name, loaded_weight in weights:
+            if "lm_head" not in name:
+                name = "model." + name
+            model_weights[name] = loaded_weight
+
+        loader.load_weights(model_weights.items())

vllm/model_executor/models/registry.py

@@ -244,6 +244,7 @@ _SPECULATIVE_DECODING_MODELS = {
     "MiMoMTPModel": ("mimo_mtp", "MiMoMTP"),
     "EAGLEModel": ("eagle", "EAGLE"),
     "EagleLlamaForCausalLM": ("llama_eagle", "EagleLlamaForCausalLM"),
+    "EagleLlama4ForCausalLM": ("llama4_eagle", "EagleLlama4ForCausalLM"),
     "EagleMiniCPMForCausalLM": ("minicpm_eagle", "EagleMiniCPMForCausalLM"),
     "Eagle3LlamaForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
     "DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),