[model_free_ptq] Enhance to work with previously quantized checkpoints like nvidia/DeepSeek-R1-NVFP4

examples/model_free_ptq/dsr1_nvfp4_fp8_block.py

@@ -0,0 +1,56 @@
+from llmcompressor import model_free_ptq
+from compressed_tensors.quantization import (
+    QuantizationScheme,
+    QuantizationArgs,
+    QuantizationStrategy,
+    QuantizationType,
+)
+
+MODEL_ID = "nvidia/DeepSeek-R1-NVFP4"
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-FP8-BLOCK"
+
+# Apply FP8-Block to the model's compatible self_attn Linear layers
+# Once quantized, the model is saved
+# using compressed-tensors to the SAVE_DIR.
+model_free_ptq(
+    model_stub=MODEL_ID,
+    save_directory=SAVE_DIR,
+    scheme=QuantizationScheme(
+        weights=QuantizationArgs(
+            num_bits=8,
+            type=QuantizationType.FLOAT,
+            strategy=QuantizationStrategy.BLOCK,
+            symmetric=True,
+            dynamic=False,
+            block_structure=[128, 128],
+        ),
+        input_activations=QuantizationArgs(
+            num_bits=8,
+            type=QuantizationType.FLOAT,
+            strategy=QuantizationStrategy.GROUP,
+            symmetric=True,
+            dynamic=True,
+            observer=None,
+            group_size=128,
+        ),
+        # TODO cannot set targets here, must be ["Linear"]
+        # targets=[
+        #     "re:.*self_attn.(o_proj|q_a_proj|q_b_proj).*"
+        # ],
+    ),
+    ignore=[
+        # NOTE: self_attn.kv_a_proj_with_mqa has incompatible shape 576x7168 with block size 128x128
+        # NOTE: self_attn.kv_b_proj is already dequantized by MLA
+        # This regex matches all strings that don't contain one of the following substrings:
+        #   - self_attn.o_proj
+        #   - self_attn.q_a_proj
+        #   - self_attn.q_b_proj
+        "re:^(?!.*self_attn.(o_proj|q_a_proj|q_b_proj)).*$"
+    ],
+    max_workers=8,
+    device="cuda:0",
+)
+
+# TODO reverse modelopt NVFP4 tensor packing order
+
+# TODO merge hf_quant_config.json with CT quantization_config in config.json