Merge pull request #33 from abhinavkulkarni/dev/more_models

awq/entry.py

@@ -73,9 +73,9 @@ def build_model_and_enc(model_path):
     # all hf model
     config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
     if "mpt" in config.__class__.__name__.lower():
-        enc = AutoTokenizer.from_pretrained(config.tokenizer_name)
+        enc = AutoTokenizer.from_pretrained(config.tokenizer_name, trust_remote_code=True)
     else:
-        enc = AutoTokenizer.from_pretrained(model_path, use_fast=False)
+        enc = AutoTokenizer.from_pretrained(model_path, use_fast=False, trust_remote_code=True)
 
     if args.load_quant:  # directly load quantized weights
         print("Loading pre-computed quantized weights...")

awq/quantize/auto_scale.py

@@ -107,11 +107,14 @@ def auto_scale_block(module, module_kwargs,
     def _search_module_scale(block, linears2scale: list, x, kwargs={}):
         # w: co, ci
         # x: n, ci
-        x = x.to(next(block.parameters()).device)
         weight = torch.cat([_m.weight for _m in linears2scale], dim=0)
         w_max = get_weight_scale(
             weight, q_group_size=q_config.get("q_group_size", -1))
+        # Clear GPU memory
+        del weight
+        torch.cuda.empty_cache()
 
+        x = x.to(next(block.parameters()).device)
         with torch.no_grad():
             org_out = block(x, **kwargs)
             if isinstance(org_out, tuple):
@@ -126,6 +129,8 @@ def auto_scale_block(module, module_kwargs,
         n_grid = 20
         history = []
 
+        # Clear GPU memory
+        torch.cuda.empty_cache()
         org_sd = {k: v.cpu() for k, v in block.state_dict().items()}
         for ratio in range(n_grid):
             ratio = ratio * 1 / n_grid

awq/quantize/pre_quant.py

@@ -135,6 +135,9 @@ def run_awq(
         # now solve for scaling and clipping
         input_feat = {k: torch.cat(v, dim=0) for k, v in input_feat.items()}
 
+        # Clear GPU memory
+        torch.cuda.empty_cache()
+
         if auto_scale:  # if it applies, we should also modify the input_feat with scales
             scales_list = auto_scale_block(
                 layer, layer_kwargs,
@@ -146,6 +149,9 @@ def run_awq(
             # append prefix to make names global
             awq_results["scale"] += append_str_prefix(scales_list, get_op_name(model, layer) + ".")
 
+        # Clear GPU memory
+        torch.cuda.empty_cache()
+        
         if mse_range:
             clip_list = auto_clip_block(layer,
                             w_bit=w_bit, q_config=q_config,