add nvfpp-1

Signed-off-by: yiliu30 <yi4.liu@intel.com>

Author: yiliu30

examples/quantization_w4a4_fp4/llama3_example.py

@@ -9,10 +9,12 @@
 MODEL_ID = "/data5/yliu7/HF_HOME/meta-llama/Llama-3.2-1B-Instruct/"
 # MODEL_ID = "meta-llama/Llama-3.3-70B-Instruct"
 scheme_name = "NVFP4"
-scheme_name = "MXFP4"
+
 # scheme_name = "MXFP8"
 # scheme_name = "FP8"
-
+scheme_name = "NVFPP_B32"
+# scheme_name = "MXFP4"
+# scheme_name = ""
 SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + f"-{scheme_name}"
 SAVE_DIR = f"/data5/yliu7/HF_HOME/{SAVE_DIR}"
 print(f"Saving to {SAVE_DIR}")
@@ -85,6 +87,7 @@ def tokenize(sample):
 input_ids = tokenizer("Hello my name is", return_tensors="pt").input_ids.to(
     model.device
 )
+print(f"=========== Starting generation =================")
 output = model.generate(input_ids, max_new_tokens=10)
 
 print(tokenizer.decode(output[0]))

src/llmcompressor/modifiers/quantization/calibration.py

@@ -14,6 +14,7 @@
     is_kv_cache_quant_scheme,
     is_mx,
     is_mxfp4,
+    use_global_scales
 )
 from compressed_tensors.utils import align_module_device, update_parameter_data
 from compressed_tensors.quantization.utils import is_kv_cache_quant_scheme
@@ -149,7 +150,10 @@ def update_weight_global_scale(module: Module):
         return
     weight_quant_args = getattr_chain(module, "quantization_scheme.weights")
     if is_mx(quantization_args=weight_quant_args):
-# MX schemes do not use global scale
+        # MX schemes do not use global scale
+        return
+    if not use_global_scales(quantization_args=weight_quant_args):
+        # global scales already in use
         return
 
     call_observer(
@@ -209,6 +213,8 @@ def calibrate_activations(module: Module, value: torch.Tensor, base_name: str):
             calculate_qparams = False
         if is_fp4(quantization_args=quantization_args):
             calculate_gparam = True
+        if not use_global_scales(quantization_args=quantization_args):
+            calculate_gparam = False
 
     call_observer(
         module=module,

src/llmcompressor/modifiers/utils/helpers.py

@@ -2,7 +2,7 @@
 
 import torch
 from compressed_tensors.quantization import QuantizationStrategy
-from compressed_tensors.quantization.utils import is_fp4, is_mxfp4
+from compressed_tensors.quantization.utils import is_fp4, is_mxfp4, use_global_scales
 from compressed_tensors.utils import align_modules, update_parameter_data
 from torch.nn import Linear, Module
 
@@ -52,6 +52,8 @@ def _valid_tensor_group_quant(layer_list: List[Linear]):
                 return False
             if not is_fp4(quantization_args=weight_quant_args):
                 return False
+            if not use_global_scales(quantization_args=weight_quant_args):
+                return False
         return True
 
     if _is_attention_module(submodule):

src/llmcompressor/transformers/compression/quantization_format.py

@@ -9,6 +9,7 @@
 )
 from compressed_tensors.quantization.utils import is_module_quantized
 from loguru import logger
+from compressed_tensors.quantization.utils.helpers import is_nvfpp_b32
 
 __all__ = ["infer_and_set_per_module_quantization_format"]
 
@@ -26,6 +27,8 @@ def _get_quant_compression_format(
     if weight_args.num_bits == 4 and weight_args.type == QuantizationType.FLOAT.value:
         if weight_args.is_mx:
              return CompressionFormat.mxfp4_pack_quantized
+        if is_nvfpp_b32(weight_args):
+            return CompressionFormat.nvfpp_b32_pack_quantized
         return CompressionFormat.nvfp4_pack_quantized
 
     if is_weight_only:  # w4a16 and w8a16