[Model] Add Gemma 2 (vllm-project#5908)

Author: WoosukKwon

docs/source/models/supported_models.rst

@@ -55,6 +55,10 @@ Alongside each architecture, we include some popular models that use it.
     - Gemma
     - :code:`google/gemma-2b`, :code:`google/gemma-7b`, etc.
     - ✅︎
+  * - :code:`Gemma2ForCausalLM`
+    - Gemma2
+    - :code:`google/gemma-2-9b`, :code:`google/gemma-2-27b`, etc.
+    - ✅︎
   * - :code:`GPT2LMHeadModel`
     - GPT-2
     - :code:`gpt2`, :code:`gpt2-xl`, etc.

requirements-common.txt

@@ -6,7 +6,7 @@ numpy < 2.0.0
 requests
 tqdm
 py-cpuinfo
-transformers >= 4.40.0  # Required for StarCoder2 & Llava, Llama 3.
+transformers >= 4.42.0  # Required for Gemma 2.
 tokenizers >= 0.19.1  # Required for Llama 3.
 fastapi
 aiohttp

vllm/config.py

@@ -14,7 +14,7 @@
 from vllm.tracing import is_otel_installed
 from vllm.transformers_utils.config import get_config, get_hf_text_config
 from vllm.utils import (cuda_device_count_stateless, get_cpu_memory, is_cpu,
-                        is_hip, is_neuron, is_tpu, is_xpu,
+                        is_hip, is_neuron, is_tpu, is_xpu, print_warning_once,
                         update_environment_variables)
 
 if TYPE_CHECKING:
@@ -141,6 +141,17 @@ def __init__(
                                     code_revision, rope_scaling, rope_theta)
         self.hf_text_config = get_hf_text_config(self.hf_config)
         self.dtype = _get_and_verify_dtype(self.hf_text_config, dtype)
+
+        if (not self.disable_sliding_window
+                and self.hf_text_config.model_type == "gemma2"
+                and self.hf_text_config.sliding_window is not None):
+            print_warning_once(
+                "Gemma 2 uses sliding window attention for every odd layer, "
+                "which is currently not supported by vLLM. Disabling sliding "
+                "window and capping the max length to the sliding window size "
+                f"({self.hf_text_config.sliding_window}).")
+            self.disable_sliding_window = True
+
         self.max_model_len = _get_and_verify_max_len(
             hf_config=self.hf_text_config,
             max_model_len=max_model_len,
@@ -257,8 +268,7 @@ def verify_with_parallel_config(
                 "BitAndBytes quantization with TP or PP is not supported yet.")
 
     def get_hf_config_sliding_window(self) -> Optional[int]:
-        """Get the sliding window size, or None if disabled.
-        """
+        """Get the sliding window size, or None if disabled."""
 
         # Some models, like Qwen2 and Qwen1.5, use `use_sliding_window` in
         # addition to sliding window size. We check if that field is present
@@ -1256,10 +1266,16 @@ def _get_and_verify_dtype(
         dtype = dtype.lower()
         if dtype == "auto":
             if config_dtype == torch.float32:
-                # Following the common practice, we use float16 for float32
-                # models.
-                logger.info("Casting torch.float32 to torch.float16.")
-                torch_dtype = torch.float16
+                if config.model_type == "gemma2":
+                    logger.info(
+                        "For Gemma 2, we downcast float32 to bfloat16 instead "
+                        "of float16 by default. Please specify `dtype` if you "
+                        "want to use float16.")
+                    torch_dtype = torch.bfloat16
+                else:
+                    # Following the common practice, we use float16 for float32
+                    # models.
+                    torch_dtype = torch.float16
             else:
                 torch_dtype = config_dtype
         else:

vllm/lora/layers.py

@@ -1069,6 +1069,10 @@ def vocab_size(self):
     def scale(self):
         return self.base_layer.scale
 
+    @property
+    def soft_cap(self):
+        return self.base_layer.soft_cap
+
     @property
     def org_vocab_size(self):
         return self.base_layer.org_vocab_size

vllm/model_executor/layers/layernorm.py

@@ -95,3 +95,49 @@ def extra_repr(self) -> str:
         s = f"hidden_size={self.weight.data.size(0)}"
         s += f", eps={self.variance_epsilon}"
         return s
+
+
+class GemmaRMSNorm(CustomOp):
+    """RMS normalization for Gemma.
+
+    Two differences from the above RMSNorm:
+        1. x * (1 + w) instead of x * w.
+        2. (x * w).to(orig_dtype) instead of x.to(orig_dtype) * w.
+    """
+
+    def __init__(
+        self,
+        hidden_size: int,
+        eps: float = 1e-6,
+    ) -> None:
+        super().__init__()
+        self.weight = nn.Parameter(torch.zeros(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward_native(
+        self,
+        x: torch.Tensor,
+        residual: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """PyTorch-native implementation equivalent to forward()."""
+        orig_dtype = x.dtype
+        if residual is not None:
+            x = x + residual
+            residual = x
+
+        x = x.float()
+        variance = x.pow(2).mean(dim=-1, keepdim=True)
+        x = x * torch.rsqrt(variance + self.variance_epsilon)
+        # Llama does x.to(float16) * w whilst Gemma is (x * w).to(float16)
+        # See https://github.com/huggingface/transformers/pull/29402
+        x = x * (1.0 + self.weight.float())
+        x = x.to(orig_dtype)
+        return x if residual is None else (x, residual)
+
+    def forward_cuda(
+        self,
+        x: torch.Tensor,
+        residual: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        # TODO(woosuk): Implement an optimized kernel for GemmaRMSNorm.
+        return self.forward_native(x, residual)

vllm/model_executor/layers/logits_processor.py

@@ -22,7 +22,8 @@ def __init__(self,
                  vocab_size: int,
                  org_vocab_size: Optional[int] = None,
                  scale: float = 1.0,
-                 logits_as_input: bool = False) -> None:
+                 logits_as_input: bool = False,
+                 soft_cap: Optional[float] = None) -> None:
         """
         Args:
             scale: A scaling factor to apply to the logits.
@@ -34,6 +35,8 @@ def __init__(self,
         self.logits_as_input = logits_as_input
         # original vocabulary size (without LoRA).
         self.org_vocab_size = org_vocab_size or vocab_size
+        # Soft cap the logits. Used in Gemma 2.
+        self.soft_cap = soft_cap
 
     def forward(
         self,
@@ -52,6 +55,11 @@ def forward(
             logits = self._get_logits(hidden_states, embedding, embedding_bias)
 
         if logits is not None:
+            if self.soft_cap is not None:
+                logits = logits / self.soft_cap
+                logits = torch.tanh(logits)
+                logits = logits * self.soft_cap
+
             if self.scale != 1.0:
                 logits *= self.scale
 

vllm/model_executor/layers/rotary_embedding.py

@@ -610,6 +610,16 @@ def forward(
         return query.flatten(-2), key.flatten(-2)
 
 
+class GemmaRotaryEmbedding(RotaryEmbedding):
+
+    def _compute_inv_freq(self, base: Union[int, float]) -> torch.Tensor:
+        # https://github.com/huggingface/transformers/blob/v4.41.2/src/transformers/models/gemma/modeling_gemma.py#L107
+        inv_freq = 1.0 / (base**(
+            torch.arange(0, self.rotary_dim, 2, dtype=torch.int64).float() /
+            self.rotary_dim))
+        return inv_freq
+
+
 _ROPE_DICT: Dict[Tuple, RotaryEmbedding] = {}
 
 

vllm/model_executor/models/__init__.py

@@ -23,6 +23,7 @@
     "DeepseekForCausalLM": ("deepseek", "DeepseekForCausalLM"),
     "FalconForCausalLM": ("falcon", "FalconForCausalLM"),
     "GemmaForCausalLM": ("gemma", "GemmaForCausalLM"),
+    "Gemma2ForCausalLM": ("gemma2", "Gemma2ForCausalLM"),
     "GPT2LMHeadModel": ("gpt2", "GPT2LMHeadModel"),
     "GPTBigCodeForCausalLM": ("gpt_bigcode", "GPTBigCodeForCausalLM"),
     "GPTJForCausalLM": ("gpt_j", "GPTJForCausalLM"),