[Model] Adds Phi-3 support (#4298)

README.md

@@ -78,6 +78,7 @@ vLLM seamlessly supports many Hugging Face models, including the following archi
 - OPT (`facebook/opt-66b`, `facebook/opt-iml-max-30b`, etc.)
 - Orion (`OrionStarAI/Orion-14B-Base`, `OrionStarAI/Orion-14B-Chat`, etc.)
 - Phi (`microsoft/phi-1_5`, `microsoft/phi-2`, etc.)
+- Phi3 (`microsoft/Phi-3-mini-4k-instruct`, `microsoft/Phi-3-mini-128k-instruct`, etc.)
 - Qwen (`Qwen/Qwen-7B`, `Qwen/Qwen-7B-Chat`, etc.)
 - Qwen2 (`Qwen/Qwen1.5-7B`, `Qwen/Qwen1.5-7B-Chat`, etc.)
 - Qwen2MoE (`Qwen/Qwen1.5-MoE-A2.7B`, `Qwen/Qwen1.5-MoE-A2.7B-Chat`, etc.)

docs/source/models/supported_models.rst

@@ -115,6 +115,10 @@ Alongside each architecture, we include some popular models that use it.
     - Phi
     - :code:`microsoft/phi-1_5`, :code:`microsoft/phi-2`, etc.
     -
+  * - :code:`Phi3ForCausalLM`
+    - Phi-3
+    - :code:`microsoft/Phi-3-mini-4k-instruct`, :code:`microsoft/Phi-3-mini-128k-instruct`, etc.
+    -
   * - :code:`QWenLMHeadModel`
     - Qwen
     - :code:`Qwen/Qwen-7B`, :code:`Qwen/Qwen-7B-Chat`, etc.

vllm/config.py

@@ -1056,7 +1056,7 @@ def _get_and_verify_max_len(
         derived_max_model_len = default_max_len
 
     rope_scaling = getattr(hf_config, "rope_scaling", None)
-    if rope_scaling is not None:
+    if rope_scaling is not None and rope_scaling["type"] != "su":
         assert "factor" in rope_scaling
         scaling_factor = rope_scaling["factor"]
         if rope_scaling["type"] == "yarn":

vllm/model_executor/layers/rotary_embedding.py

@@ -338,6 +338,114 @@ class YaRNScalingRotaryEmbedding(RotaryEmbedding):
         return cache
 
 
+class Phi3SuScaledRotaryEmbedding(nn.Module):
+    """Phi3 family of models scaled rotary embedding.
+
+    Based on the original RotaryEmbedding implementation.
+    """
+
+    def __init__(
+        self,
+        head_size: int,
+        rotary_dim: int,
+        max_position_embeddings: int,
+        original_max_position_embeddings: int,
+        base: int,
+        is_neox_style: bool,
+        short_factor: List[float],
+        long_factor: List[float],
+        short_mscale: float = 1.1,
+        long_mscale: float = 1.225,
+    ):
+        super().__init__()
+
+        if rotary_dim != head_size:
+            raise ValueError(
+                f"`Phi3SuScaledRotaryEmbedding` does not support rotary_dim != \
+                    head_size ({rotary_dim}!={head_size}).")
+        if is_neox_style is False:
+            raise ValueError(
+                "`Phi3SuScaledRotaryEmbedding` only supports neox_style.")
+
+        self.head_size = head_size
+        self.max_position_embeddings = max_position_embeddings
+        self.original_max_position_embeddings = original_max_position_embeddings
+        self.base = base
+        self.short_factor = short_factor
+        self.long_factor = long_factor
+        self.short_mscale = short_mscale
+        self.long_mscale = long_mscale
+
+        short_cache = self._compute_cos_sin_cache(
+            original_max_position_embeddings, short_factor, short_mscale)
+        short_cache = short_cache.to(torch.get_default_dtype())
+        self.register_buffer("short_cos_sin_cache",
+                             short_cache,
+                             persistent=False)
+
+        long_cache = self._compute_cos_sin_cache(max_position_embeddings,
+                                                 long_factor, long_mscale)
+        long_cache = long_cache.to(torch.get_default_dtype())
+        self.register_buffer("long_cos_sin_cache",
+                             long_cache,
+                             persistent=False)
+
+        long_short_cache = torch.cat(
+            [self.short_cos_sin_cache, self.long_cos_sin_cache], dim=0)
+        self.register_buffer("long_short_cos_sin_cache",
+                             long_short_cache,
+                             persistent=False)
+
+    def _compute_inv_freq(self, rescale_factors: List[float]) -> torch.Tensor:
+        rescale_factors = torch.tensor(rescale_factors, dtype=torch.float32)
+        inv_freq = 1.0 / (rescale_factors * (self.base**(torch.arange(
+            0, self.head_size, 2, dtype=torch.float) / self.head_size)))
+        return inv_freq
+
+    def _compute_cos_sin_cache(
+        self,
+        max_position_embeddings: int,
+        rescale_factors: List[float],
+        mscale: float,
+    ) -> torch.Tensor:
+        inv_freq = self._compute_inv_freq(rescale_factors)
+        t = torch.arange(max_position_embeddings, dtype=torch.float)
+        freqs = torch.einsum("i,j -> ij", t, inv_freq)
+        cos = freqs.cos() * mscale
+        sin = freqs.sin() * mscale
+        cache = torch.cat((cos, sin), dim=-1)
+        return cache
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        offsets: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        query = query.view(*query.shape[:-1], -1, self.head_size)
+        key = key.view(*key.shape[:-1], -1, self.head_size)
+
+        k = self.original_max_position_embeddings
+        long_prompt_offset = (torch.any(positions > k).float() *
+                              torch.full_like(positions, k)).long()
+        idx = (torch.add(positions, long_prompt_offset)
+               if long_prompt_offset is not None else positions)
+        self.long_short_cos_sin_cache = self.long_short_cos_sin_cache.to(
+            idx.device)
+        idx = torch.add(idx, offsets) if offsets is not None else idx
+        cos_sin = torch.index_select(self.long_short_cos_sin_cache, 0, idx)
+
+        cos, sin = cos_sin.chunk(2, dim=-1)
+        cos = cos.repeat(1, 2).unsqueeze(-2)
+        sin = sin.repeat(1, 2).unsqueeze(-2)
+
+        query = query * cos + _rotate_neox(query) * sin
+        key = key * cos + _rotate_neox(key) * sin
+
+        return query.flatten(-2), key.flatten(-2)
+
+
 _ROPE_DICT: Dict[Tuple, RotaryEmbedding] = {}
 
 
@@ -349,16 +457,25 @@ def get_rope(
     is_neox_style: bool = True,
     rope_scaling: Optional[Dict[str, Any]] = None,
 ) -> RotaryEmbedding:
+    if rope_scaling is not None:
+        # Transforms every value that is a list into a tuple for caching calls
+        rope_scaling_tuple = {
+            k: tuple(v) if isinstance(v, list) else v
+            for k, v in rope_scaling.items()
+        }
+        rope_scaling_args = tuple(rope_scaling_tuple.items())
+    else:
+        rope_scaling_args = None
     key = (head_size, rotary_dim, max_position, base, is_neox_style,
-           tuple(rope_scaling.items()) if rope_scaling is not None else None)
+           rope_scaling_args)
     if key in _ROPE_DICT:
         return _ROPE_DICT[key]
-
     if rope_scaling is None:
         rotary_emb = RotaryEmbedding(head_size, rotary_dim, max_position, base,
                                      is_neox_style)
     else:
         scaling_type = rope_scaling["type"]
+        if scaling_type != "su":
             scaling_factor = rope_scaling["factor"]
         if scaling_type == "linear":
             rotary_emb = LinearScalingRotaryEmbedding(head_size, rotary_dim,
@@ -383,6 +500,19 @@ def get_rope(
                                                     base, is_neox_style,
                                                     scaling_factor,
                                                     **extra_kwargs)
+        elif scaling_type == "su":
+            short_factor = rope_scaling["short_factor"]
+            long_factor = rope_scaling["long_factor"]
+            original_max_position = rope_scaling[
+                "original_max_position_embeddings"]
+            extra_kwargs = {
+                k: v
+                for k, v in rope_scaling.items()
+                if k in ("short_mscale", "long_mscale")
+            }
+            rotary_emb = Phi3SuScaledRotaryEmbedding(
+                head_size, rotary_dim, max_position, original_max_position,
+                base, is_neox_style, short_factor, long_factor, **extra_kwargs)
         else:
             raise ValueError(f"Unknown RoPE scaling type {scaling_type}")
     _ROPE_DICT[key] = rotary_emb

vllm/model_executor/models/__init__.py

@@ -46,6 +46,7 @@ _MODELS = {
     "OPTForCausalLM": ("opt", "OPTForCausalLM"),
     "OrionForCausalLM": ("orion", "OrionForCausalLM"),
     "PhiForCausalLM": ("phi", "PhiForCausalLM"),
+    "Phi3ForCausalLM": ("llama", "LlamaForCausalLM"),
     "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
     "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
     "Qwen2MoeForCausalLM": ("qwen2_moe", "Qwen2MoeForCausalLM"),

vllm/model_executor/models/llama.py

@@ -180,6 +180,10 @@ class LlamaDecoderLayer(nn.Module):
         self.hidden_size = config.hidden_size
         rope_theta = getattr(config, "rope_theta", 10000)
         rope_scaling = getattr(config, "rope_scaling", None)
+        if rope_scaling is not None and getattr(
+                config, "original_max_position_embeddings", None):
+            rope_scaling["original_max_position_embeddings"] = (
+                config.original_max_position_embeddings)
         max_position_embeddings = getattr(config, "max_position_embeddings",
                                           8192)
         sliding_window = getattr(config, "sliding_window", None)
@@ -378,11 +382,11 @@ class LlamaForCausalLM(nn.Module):
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [
             # (param_name, shard_name, shard_id)
-            ("qkv_proj", "q_proj", "q"),
+            (".qkv_proj", ".q_proj", "q"),
-            ("qkv_proj", "k_proj", "k"),
+            (".qkv_proj", ".k_proj", "k"),
-            ("qkv_proj", "v_proj", "v"),
+            (".qkv_proj", ".v_proj", "v"),
-            ("gate_up_proj", "gate_proj", 0),
+            (".gate_up_proj", ".gate_proj", 0),
-            ("gate_up_proj", "up_proj", 1),
+            (".gate_up_proj", ".up_proj", 1),
         ]
         params_dict = dict(self.named_parameters())
         for name, loaded_weight in weights: