Merge pull request #77 from s4rduk4r/offloading

Offloading to cpu and disk

awq/models/auto.py

@@ -41,11 +41,13 @@ class AutoAWQForCausalLM:
     @classmethod
     def from_quantized(self, quant_path, quant_filename='', max_new_tokens=None,
                        trust_remote_code=True, fuse_layers=True,
-                       batch_size=1, safetensors=False) -> BaseAWQForCausalLM:
+                       batch_size=1, safetensors=False,
+                       max_memory=None, offload_folder=None) -> BaseAWQForCausalLM:
         os.environ["AWQ_BATCH_SIZE"] = str(batch_size)
         model_type = check_and_get_model_type(quant_path, trust_remote_code)
 
         return AWQ_CAUSAL_LM_MODEL_MAP[model_type].from_quantized(
             quant_path, model_type, quant_filename, max_new_tokens, trust_remote_code=trust_remote_code, 
-            fuse_layers=fuse_layers, safetensors=safetensors
-        )
\ No newline at end of file
+            fuse_layers=fuse_layers, safetensors=safetensors, 
+            max_memory=max_memory, offload_folder=offload_folder
+        )

awq/models/base.py

@@ -133,7 +133,8 @@ class BaseAWQForCausalLM(nn.Module):
     def from_quantized(self, model_path, model_type, model_filename='', 
                              max_new_tokens=None, torch_dtype=torch.float16, 
                              trust_remote_code=True, safetensors=False, is_quantized=True, 
-                             fuse_layers=False, version='GEMM'):
+                             fuse_layers=False, version='GEMM',
+                             max_memory=None, offload_folder=None):
         # [STEP 1-2] Load weights path and configs
         model_weights_path, config, quant_config = self._load_config(
             self, model_path, model_filename, safetensors, version, 
@@ -153,6 +154,7 @@ class BaseAWQForCausalLM(nn.Module):
         device_map = infer_auto_device_map(
             model,
             no_split_module_classes=[self.layer_type], 
+            max_memory=max_memory,
             dtype=torch_dtype
         )
 
@@ -160,15 +162,24 @@ class BaseAWQForCausalLM(nn.Module):
         load_checkpoint_in_model(
             model,
             checkpoint=model_weights_path,
-            device_map=device_map
+            device_map=device_map,
+            offload_folder=offload_folder,
+            dtype=torch_dtype
         )
         
         # Dispath to devices
-        model = simple_dispatch_model(model, device_map)
-        
         if fuse_layers:
             self.fuse_layers(model, quant_config)
 
+        # Offloading dispatch
+        from accelerate import dispatch_model
+        model = dispatch_model(
+            model,
+            device_map=device_map,
+            offload_dir=offload_folder
+        )
+        
+
         return self(model, model_type, is_quantized=is_quantized, quant_config=quant_config)
 
     def _load_config(self, model_path, model_filename, safetensors=False, 

awq/modules/fused/attn.py

@@ -36,7 +36,7 @@ def apply_rotary_emb(
     xk_ = torch.view_as_complex(
         xk.float().reshape(*xk.shape[:-1], 2, -1).transpose(-2, -1).contiguous()
     )
-    freqs_cis = reshape_for_broadcast(freqs_cis, xq_)
+    freqs_cis = reshape_for_broadcast(freqs_cis, xq_).to(xq_.device)
     xq_out = torch.view_as_real(xq_ * freqs_cis).transpose(-2, -1).flatten(3)
     xk_out = torch.view_as_real(xk_ * freqs_cis).transpose(-2, -1).flatten(3)
     return xq_out.type_as(xq), xk_out.type_as(xk)