Fix AutoModel can't load gptq model due to module prefix mismatch vs AutoModelForCausalLM (#2146)

* fix the issue of AutoModel failing to load the gptq model.

* clear

* update comments

---------

Co-authored-by: Qubitium-ModelCloud <qubitium@modelcloud.ai>

Author: LRL-ModelCloud

optimum/gptq/constants.py

@@ -18,6 +18,10 @@
     "model.decoder.layers",
     "gpt_neox.layers",
     "model.layers",
+    # modules loaded by AutoModel vs AutoModelForCausalLM have different prefixes
+    "h",
+    "decoder.layers",
+    "layers",
 ]
 
 GPTQ_CONFIG = "quantize_config.json"

optimum/gptq/utils.py

@@ -72,7 +72,7 @@ def get_block_name_with_pattern(model: nn.Module):
     modules_names = [n for n, _ in model.named_modules()]
     for pattern_candidate in BLOCK_PATTERNS:
         pattern_candidate = pattern_candidate
-        if any(pattern_candidate in name for name in modules_names):
+        if any(name.startswith(pattern_candidate) for name in modules_names):
             return pattern_candidate
     raise ValueError("Block pattern could not be match. Pass `block_name_to_quantize` argument in `quantize_model`")