[lut-b] data-free round-to-nearest example

examples/quantization_lutb/llama3_example.py

@@ -0,0 +1,37 @@
+from compressed_tensors.offload import dispatch_model
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from llmcompressor import oneshot
+from llmcompressor.modifiers.quantization import QuantizationModifier
+
+MODEL_ID = "meta-llama/Meta-Llama-3-8B-Instruct"
+
+# Load model.
+model = AutoModelForCausalLM.from_pretrained(MODEL_ID)
+tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
+
+# Configure the quantization algorithm and scheme.
+# In this case, we:
+#   * quantize only the mlp layer weights
+recipe = QuantizationModifier(
+    targets="re:.*layers.*mlp.*_proj$", scheme="LUTB", ignore=["lm_head"]
+)
+
+# Apply quantization.
+oneshot(model=model, recipe=recipe)
+
+print("\n\n")
+print("========== SAMPLE GENERATION ==============")
+dispatch_model(model)
+input_ids = tokenizer("Hello my name is", return_tensors="pt").input_ids.to(
+    model.device
+)
+output = model.generate(input_ids, max_new_tokens=100)
+print(tokenizer.decode(output[0]))
+print("==========================================\n\n")
+
+
+# Save to disk in compressed-tensors format.
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-LUTB"
+model.save_pretrained(SAVE_DIR, save_compressed=True)
+tokenizer.save_pretrained(SAVE_DIR)

src/llmcompressor/modifiers/quantization/calibration.py

@@ -77,7 +77,7 @@ def initialize_observer(
             log_once=True,
         )
 
-    if args is not None and args.dynamic is not True:
+    if args is not None and args.dynamic is not True and observer is not None:
         observer = Observer.load_from_registry(observer, base_name=base_name, args=args)
         module.register_module(f"{base_name}_observer", observer)
         observer.attach(module)