[V1][Spec Decode] Share input embedding of target model with EAGLE draft model to free ~1GB for llama 3 model (#17326)

Co-authored-by: root <root@ekagra-8xh100.us-east5-a.c.serving-efficiency-poc.internal>
Co-authored-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>

Author: 3 people

examples/offline_inference/eagle.py

@@ -105,6 +105,13 @@ def main():
     outputs = llm.generate(prompt_token_ids=prompt_ids,
                            sampling_params=sampling_params)
 
+    # print the generated text
+    for output in outputs:
+        print("-" * 50)
+        print(f"prompt: {output.prompt}")
+        print(f"generated text: {output.outputs[0].text}")
+        print("-" * 50)
+
     if not hasattr(outputs, "metrics") or outputs.metrics is None:
         return
 

vllm/model_executor/models/llama_eagle.py

@@ -8,6 +8,7 @@
 
 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.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.vocab_parallel_embedding import (
@@ -52,11 +53,15 @@ def __init__(
         self.config = vllm_config. \
             speculative_config.draft_model_config.hf_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"),
-        )
+
+        # if PP disabled then draft will share embed with target
+        if get_pp_group().world_size > 1:
+            self.embed_tokens = VocabParallelEmbedding(
+                self.config.vocab_size,
+                self.config.hidden_size,
+                prefix=maybe_prefix(prefix, "embed_tokens"),
+            )
+
         self.layers = nn.ModuleList([
             LlamaDecoderLayer(
                 self.config,
@@ -109,6 +114,12 @@ def load_weights(self, weights: Iterable[Tuple[str,
                 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)
@@ -142,14 +153,12 @@ def forward(
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         loader = AutoWeightsLoader(
             self,
-            skip_prefixes=(["lm_head."]
-                           if self.config.tie_word_embeddings else None),
+            skip_prefixes=None,
         )
 
         model_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())
+        return loader.load_weights(model_weights.items())

vllm/model_executor/models/llama_eagle3.py

@@ -8,6 +8,7 @@
 
 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.linear import QKVParallelLinear
@@ -91,11 +92,15 @@ def __init__(
         self.config = vllm_config. \
             speculative_config.draft_model_config.hf_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"),
-        )
+
+        # if PP disabled then draft will share embed with target
+        if get_pp_group().world_size > 1:
+            self.embed_tokens = VocabParallelEmbedding(
+                self.config.vocab_size,
+                self.config.hidden_size,
+                prefix=maybe_prefix(prefix, "embed_tokens"),
+            )
+
         self.layers = nn.ModuleList([
             LlamaDecoderLayer(
                 self.config,

vllm/v1/spec_decode/eagle.py

@@ -5,6 +5,7 @@
 from vllm.attention.layer import Attention
 from vllm.config import (CompilationLevel, VllmConfig,
                          get_layers_from_vllm_config, set_current_vllm_config)
+from vllm.distributed.parallel_state import get_pp_group
 from vllm.forward_context import set_forward_context
 from vllm.logger import init_logger
 from vllm.model_executor.model_loader import get_model_loader
@@ -306,12 +307,30 @@ def load_model(self, target_model: nn.Module) -> None:
         self.attn_layer_name = next(iter(draft_attn_layer_names))
         loaded_weights = self.model.load_weights(
             loader.get_all_weights(draft_model_config, self.model))
-        if self.vllm_config.speculative_config.method == "eagle3":
-            if "model.embed_tokens.weight" not in loaded_weights:
-                logger.info(
-                    "Loading EAGLE embedding weights from the target model.")
-                self.model.model.embed_tokens = target_model.model.embed_tokens
+
+        # share embed_tokens with the target model if needed
+        if get_pp_group().world_size == 1:
+            assert "model.embed_tokens.weight" not in loaded_weights, \
+            "For PP = 1, Eagle draft should share embed with target model"
+            logger.info(
+                "The EAGLE head shares the same vocab embedding" \
+                " with the target model."
+            )
+            self.model.model.embed_tokens = target_model.model.embed_tokens
         else:
+            assert "model.embed_tokens.weight" in loaded_weights, \
+            "For PP > 1, Eagle draft checkpoint should its own copy of "
+            " the model.embed_tokens.weight"
+            logger.info(
+                "Since PP > 1, the EAGLE head loaded its own vocab embedding" \
+                " weights instead of sharing them with the target model."
+            )
+
+        # share lm_head with the target model if needed
+        # some model definition do not define lm_head explicitly
+        # and reuse embed_tokens for lm_head, e.g., CohereForCausalLM
+        if self.vllm_config.speculative_config.method != "eagle3" and \
+                hasattr(target_model, "lm_head"):
             logger.info("Loading EAGLE LM head weights from the target model.")
             self.model.lm_head = target_model.lm_head