[MXFP4] Support

experimental/README.md

@@ -0,0 +1,3 @@
+# Experimental Features
+
+This folder aims to highlights features that are still a work-in-progress or potentially supported in LLM Compressor and / or Compressed-Tensors but missing support in vLLM.

experimental/mxfp4/llama3_mxfp4.py

@@ -0,0 +1,35 @@
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from llmcompressor import oneshot
+from llmcompressor.modifiers.quantization import QuantizationModifier
+from llmcompressor.utils import dispatch_for_generation
+
+MODEL_ID = "meta-llama/Meta-Llama-3-8B-Instruct"
+
+# Load model.
+model = AutoModelForCausalLM.from_pretrained(MODEL_ID, torch_dtype="auto")
+tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
+
+# Configure the quantization algorithm and scheme.
+# In this case, we:
+#   * quantize the weights to fp4 with per group 32 via ptq
+recipe = QuantizationModifier(targets="Linear", scheme="MXFP4", ignore=["lm_head"])
+
+# Apply quantization.
+oneshot(model=model, recipe=recipe)
+
+print("\n\n")
+print("========== SAMPLE GENERATION ==============")
+dispatch_for_generation(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] + "-MXFP4"
+model.save_pretrained(SAVE_DIR, save_compressed=True)
+tokenizer.save_pretrained(SAVE_DIR)