Add activation registry (#126)

Author: WoosukKwon

cacheflow/entrypoints/llm.py

@@ -61,7 +61,7 @@ def generate(
         while self.llm_server.has_unfinished_requests():
             step_outputs = self.llm_server.step()
             for output in step_outputs:
-                if output.done:
+                if output.finished():
                     outputs.append(output)
                     if use_tqdm:
                         pbar.update(1)

cacheflow/model_executor/layers/activation.py

@@ -4,6 +4,21 @@
 
 from cacheflow import activation_ops
 
+_ACTIVATION_REGISTRY = {
+    "gelu": nn.GELU(),
+    "gelu_new": nn.GELU(approximate="tanh"),   # NOTE: This may introduce small rounding errors.
+    "gelu_fast": nn.GELU(approximate="tanh"),  # NOTE: This may introduce small rounding errors.
+    "relu": nn.ReLU(),
+}
+
+
+def get_act_fn(act_fn: str) -> nn.Module:
+    """Get an activation function by name."""
+    act_fn = act_fn.lower()
+    if act_fn in _ACTIVATION_REGISTRY:
+        return _ACTIVATION_REGISTRY[act_fn]
+    raise ValueError(f"Activation function {act_fn!r} is not supported.")
+
 
 class SiluAndMul(nn.Module):
     """An activation function for SwiGLU.

cacheflow/model_executor/models/gpt2.py

@@ -27,6 +27,7 @@
 from transformers import GPT2Config
 
 from cacheflow.model_executor.input_metadata import InputMetadata
+from cacheflow.model_executor.layers.activation import get_act_fn
 from cacheflow.model_executor.layers.attention import GPTCacheFlowAttention
 from cacheflow.model_executor.layers.sampler import Sampler
 from cacheflow.model_executor.weight_utils import (hf_model_weights_iterator,
@@ -92,12 +93,7 @@ def __init__(
         self.c_proj = RowParallelLinear(intermediate_size, hidden_size,
                                         bias=True, input_is_parallel=True,
                                         perform_initialization=False)
-
-        act_fn = config.activation_function
-        if act_fn != "gelu_new":
-            raise ValueError(f"Unsupported activation: {act_fn}. "
-                             "GPT-2 only supports gelu_new for now.")
-        self.act = torch.nn.GELU(approximate="tanh")
+        self.act = get_act_fn(config.activation_function)
 
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
         hidden_states, _ = self.c_fc(hidden_states)

cacheflow/model_executor/models/gpt_neox.py

@@ -26,6 +26,7 @@
 from transformers import GPTNeoXConfig
 
 from cacheflow.model_executor.input_metadata import InputMetadata
+from cacheflow.model_executor.layers.activation import get_act_fn
 from cacheflow.model_executor.layers.attention import GPTNeoXCacheFlowAttention
 from cacheflow.model_executor.layers.sampler import Sampler
 from cacheflow.model_executor.weight_utils import (hf_model_weights_iterator,
@@ -94,10 +95,7 @@ def __init__(self, config: GPTNeoXConfig):
         self.dense_4h_to_h = RowParallelLinear(config.intermediate_size, config.hidden_size,
                                                input_is_parallel=True,
                                                perform_initialization=False)
-        if config.hidden_act != 'gelu':
-            raise ValueError(f'Unsupported activation: {config.hidden_act}. '
-                             'Only gelu is supported for now.')
-        self.act = torch.nn.GELU()
+        self.act = get_act_fn(config.hidden_act)
 
     def forward(self, hidden_states):
         hidden_states, _ = self.dense_h_to_4h(hidden_states)

cacheflow/model_executor/models/opt.py

@@ -26,6 +26,7 @@
 from transformers import OPTConfig
 
 from cacheflow.model_executor.input_metadata import InputMetadata
+from cacheflow.model_executor.layers.activation import get_act_fn
 from cacheflow.model_executor.layers.attention import GPTCacheFlowAttention
 from cacheflow.model_executor.layers.sampler import Sampler
 from cacheflow.model_executor.weight_utils import (hf_model_weights_iterator,
@@ -105,8 +106,7 @@ def __init__(self, config: OPTConfig):
             bias=config.enable_bias,
         )
         self.do_layer_norm_before = config.do_layer_norm_before
-        assert config.activation_function == 'relu'
-        self.activation_fn = nn.ReLU()
+        self.activation_fn = get_act_fn(config.activation_function)
 
         self.self_attn_layer_norm = nn.LayerNorm(
             self.embed_dim, elementwise_affine=config.layer_norm_elementwise_affine)