fix gptq models conversion (openvinotoolkit#104)

gklodkox · Dec 22, 2023 · bf42971 · bf42971
1 parent 6787f31
commit bf42971
Show file tree

Hide file tree

Showing 2 changed files with 4 additions and 3 deletions.
diff --git a/llm_bench/python/convert.py b/llm_bench/python/convert.py
@@ -810,7 +810,7 @@ def ts_patched_forward(
         args.compress_weights and not is_torch_compression(args) and is_ov_model_provided(args.model_id, args.output_dir, args.precision)
     )
     gptq_applied = is_gptq(config)
-    precision = precision if not gptq_applied else GPTQ_DIR.format(args.precision)
+    precision = precision if not gptq_applied else GPTQ_DIR.format(presision=args.precision)
     if post_init is not None:
         model_kwargs = {"torch_dtype": torch.float32}
     pt_model = None
@@ -885,7 +885,7 @@ def convert_to_ov(pt_model, tok, out_path, compress_to_fp16=False):
     )
     compress_to_fp16 = is_fp16(args)
     gptq_applied = is_gptq(config)
-    precision = precision if not gptq_applied else GPTQ_DIR.format(args.precision)
+    precision = precision if not gptq_applied else GPTQ_DIR.format(precision=args.precision)
     tokenizer_id = args.tokenizer_id or args.model_id
     tok = AutoTokenizer.from_pretrained(tokenizer_id, trust_remote_code=True)
     ov_out_path = Path(args.output_dir) / PYTORCH_DIR / OV_DIR / precision
@@ -981,7 +981,7 @@ def convert_to_ov(pt_model, tok, out_path, compress_to_fp16=False):
     tokenizer_id = args.tokenizer_id or args.model_id
     tok = AutoTokenizer.from_pretrained(tokenizer_id, trust_remote_code=True)
     gptq_applied = is_gptq(config)
-    precision = precision if not gptq_applied else GPTQ_DIR.format(args.precision)
+    precision = precision if not gptq_applied else GPTQ_DIR.format(precision=args.precision)
     if post_init is not None:
         model_kwargs = {"torch_dtype": torch.float32}
     pt_model = None

diff --git a/llm_bench/python/utils/conversion_utils/helpers.py b/llm_bench/python/utils/conversion_utils/helpers.py
@@ -41,6 +41,7 @@ def is_ov_model_provided(model_id, model_dir, precision, model_name="openvino_mo
         model_dirs.append(Path(model_id) / precision)
         model_dirs.append(Path(model_id) / OV_DIR / precision)
         model_dirs.append(Path(model_id) / PYTORCH_DIR / OV_DIR / precision)
+    model_dir = Path(model_dir)
     model_dirs.append(model_dir)
     model_dirs.append(model_dir / precision)
     model_dirs.append(model_dir / OV_DIR / precision)