Remove ScaledActivation for AWQ (#10057)

Signed-off-by: mgoin <michael@neuralmagic.com>

vllm/model_executor/layers/quantization/tpu_int8.py

@@ -50,9 +50,6 @@ class Int8TpuConfig(QuantizationConfig):
             return TPUInt8LinearMethod(self)
         return None
 
-    def get_scaled_act_names(self) -> List[str]:
-        return []
-
 
 class TPUInt8LinearMethod(LinearMethodBase):
     """Int8 Linear method for TPU Quant. """

vllm/model_executor/models/bart.py

@@ -393,8 +393,7 @@ class BartEncoderLayer(nn.Module):
             cache_config=cache_config,
             quant_config=quant_config)
         self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
-        self.activation_fn = get_act_fn(config.activation_function,
-                                        quant_config)
+        self.activation_fn = get_act_fn(config.activation_function)
 
         ffn_hidden_size = self.embed_dim
         ffn_intermediate_size = config.encoder_ffn_dim
@@ -405,7 +404,7 @@ class BartEncoderLayer(nn.Module):
             bias=ffn_has_bias,
             quant_config=quant_config,
         )
-        self.act = get_act_fn("gelu", quant_config, ffn_intermediate_size)
+        self.act = get_act_fn("gelu")
         self.fc2 = RowParallelLinear(
             ffn_intermediate_size,
             ffn_hidden_size,
@@ -473,8 +472,7 @@ class BartDecoderLayer(nn.Module):
             config=config,
             cache_config=cache_config,
             quant_config=quant_config)
-        self.activation_fn = get_act_fn(config.activation_function,
-                                        quant_config)
+        self.activation_fn = get_act_fn(config.activation_function)
 
         self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
         '''

vllm/model_executor/models/bloom.py

@@ -146,7 +146,7 @@ class BloomMLP(nn.Module):
             4 * hidden_size,
             quant_config=quant_config,
         )
-        self.gelu_impl = get_act_fn("gelu", quant_config, 4 * hidden_size)
+        self.gelu_impl = get_act_fn("gelu")
         self.dense_4h_to_h = RowParallelLinear(
             4 * hidden_size,
             hidden_size,

vllm/model_executor/models/falcon.py

@@ -212,7 +212,7 @@ class FalconMLP(nn.Module):
                                                   bias=config.bias,
                                                   skip_bias_add=True,
                                                   quant_config=quant_config)
-        self.act = get_act_fn("gelu", quant_config, 4 * hidden_size)
+        self.act = get_act_fn("gelu")
         self.reduce_row_parallel_results = not (config.new_decoder_architecture
                                                 or config.parallel_attn)
         self.dense_4h_to_h = RowParallelLinear(

vllm/model_executor/models/gpt2.py

@@ -123,8 +123,7 @@ class GPT2MLP(nn.Module):
             quant_config=quant_config,
             prefix=f"{prefix}.c_proj",
         )
-        self.act = get_act_fn(config.activation_function, quant_config,
-                              intermediate_size)
+        self.act = get_act_fn(config.activation_function)
 
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
         hidden_states, _ = self.c_fc(hidden_states)

vllm/model_executor/models/gpt_bigcode.py

@@ -135,8 +135,7 @@ class GPTBigMLP(nn.Module):
             bias=True,
             quant_config=quant_config,
         )
-        self.act = get_act_fn(config.activation_function, quant_config,
-                              intermediate_size)
+        self.act = get_act_fn(config.activation_function)
 
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
         hidden_states, _ = self.c_fc(hidden_states)

vllm/model_executor/models/gpt_j.py

@@ -130,8 +130,7 @@ class GPTJMLP(nn.Module):
             hidden_size,
             quant_config=quant_config,
         )
-        self.act = get_act_fn(config.activation_function, quant_config,
-                              intermediate_size)
+        self.act = get_act_fn(config.activation_function)
 
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
         hidden_states, _ = self.fc_in(hidden_states)

vllm/model_executor/models/gpt_neox.py

@@ -128,8 +128,7 @@ class GPTNeoXMLP(nn.Module):
             config.hidden_size,
             quant_config=quant_config,
         )
-        self.act = get_act_fn(config.hidden_act, quant_config,
-                              config.intermediate_size)
+        self.act = get_act_fn(config.hidden_act)
 
     def forward(self, hidden_states):
         hidden_states, _ = self.dense_h_to_4h(hidden_states)

vllm/model_executor/models/mpt.py

@@ -153,7 +153,7 @@ class MPTMLP(nn.Module):
             bias=not config.no_bias,
             quant_config=quant_config,
         )
-        self.act = get_act_fn("gelu", quant_config, intermediate_size)
+        self.act = get_act_fn("gelu")
         self.down_proj = RowParallelLinear(
             intermediate_size,
             hidden_size,

vllm/model_executor/models/opt.py

@@ -147,8 +147,7 @@ class OPTDecoderLayer(nn.Module):
             quant_config=quant_config,
             prefix=f"{prefix}.fc1",
         )
-        self.activation_fn = get_act_fn(config.activation_function,
-                                        quant_config, config.ffn_dim)
+        self.activation_fn = get_act_fn(config.activation_function)
         self.fc2 = RowParallelLinear(
             config.ffn_dim,
             self.embed_dim,

vllm/model_executor/models/persimmon.py

@@ -60,7 +60,7 @@ class PersimmonMLP(nn.Module):
         self.dense_4h_to_h = RowParallelLinear(config.intermediate_size,
                                                config.hidden_size,
                                                quant_config=quant_config)
-        self.act = get_act_fn(config.hidden_act, quant_config)
+        self.act = get_act_fn(config.hidden_act)
 
     def forward(self, hidden_states) -> torch.Tensor:
         hidden_states, _ = self.dense_h_to_4h(hidden_states)

vllm/model_executor/models/phi.py

@@ -152,7 +152,7 @@ class PhiMLP(nn.Module):
             config.hidden_size,
             quant_config=quant_config,
         )
-        self.act = get_act_fn(config.hidden_act, quant_config, n_inner)
+        self.act = get_act_fn(config.hidden_act)
 
     def forward(self, hidden_states):
         hidden_states, _ = self.fc1(hidden_states)

vllm/model_executor/models/qwen.py

@@ -203,7 +203,7 @@ class QwenVMLP(nn.Module):
                                          intermediate_size,
                                          bias=True,
                                          quant_config=quant_config)
-        self.act_fn = get_act_fn("gelu", quant_config, intermediate_size)
+        self.act_fn = get_act_fn("gelu")
         self.c_proj = RowParallelLinear(
             intermediate_size,
             hidden_size,

vllm/model_executor/models/starcoder2.py

@@ -139,8 +139,7 @@ class Starcoder2MLP(nn.Module):
             bias=config.use_bias,
             quant_config=quant_config,
         )
-        self.act = get_act_fn(config.hidden_act, quant_config,
-                              config.intermediate_size)
+        self.act = get_act_fn(config.hidden_act)
 
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
         hidden_states, _ = self.c_fc(hidden_states)