From 367ed705ab9cfe0853bf3186972ed8ac0b669eae Mon Sep 17 00:00:00 2001 From: marauder-actual Date: Mon, 1 Jun 2026 04:23:28 +0200 Subject: [PATCH] fix: convert chat messages to text for AutoRound calibration --- quantize_awq.py | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) diff --git a/quantize_awq.py b/quantize_awq.py index 14a18e5..746c43d 100644 --- a/quantize_awq.py +++ b/quantize_awq.py @@ -23,10 +23,14 @@ model = AutoModelForCausalLM.from_pretrained( MODEL, torch_dtype=torch.float16, device_map="auto" ) -# Domain calibration from training data +# Domain calibration from training data (chat format → raw text) print("Loading calibration data...") with open(CALIB_DATA) as f: - calib = [json.loads(line)["text"] for line in f] + calib = [] + for line in f: + msgs = json.loads(line)["messages"] + text = tokenizer.apply_chat_template(msgs, tokenize=False) + calib.append(text) print(f"Calibration samples: {len(calib)}")