[Misc] Add compressed-tensors NVFP4A16 emulation support (#17914)

Signed-off-by: Dipika Sikka <dipikasikka1@gmail.com>
Signed-off-by: Dipika <dipikasikka1@gmail.com>

tests/quantization/test_compressed_tensors.py

@@ -13,9 +13,9 @@ from compressed_tensors.quantization import QuantizationType
 from tests.models.utils import check_logprobs_close
 from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors import (  # noqa: E501
     CompressedTensors24, CompressedTensorsLinearMethod,
-    CompressedTensorsW4A16Sparse24, CompressedTensorsW8A8Fp8,
-    CompressedTensorsW8A8Int8, CompressedTensorsW8A16Fp8,
-    CompressedTensorsWNA16)
+    CompressedTensorsW4A16Fp4, CompressedTensorsW4A16Sparse24,
+    CompressedTensorsW8A8Fp8, CompressedTensorsW8A8Int8,
+    CompressedTensorsW8A16Fp8, CompressedTensorsWNA16)
 from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
     sparse_cutlass_supported)
 from vllm.platforms import current_platform
@@ -648,3 +648,23 @@ def test_compressed_tensors_2of4_sparse_compressed(vllm_runner, args_2of4):
         output = llm.generate_greedy("Hello my name is", max_tokens=20)
         print(output)
         assert output
+
+
+def test_compressed_tensors_nvfp4a16(vllm_runner):
+    # run weight only example
+    model = "nm-testing/TinyLlama-1.1B-Chat-v1.0-FP4"
+    with vllm_runner(model, enforce_eager=True) as llm:
+
+        def check_model(model):
+            layer = model.model.layers[0]
+
+            qkv_proj = layer.self_attn.qkv_proj
+            assert isinstance(qkv_proj.quant_method,
+                              CompressedTensorsLinearMethod)
+            assert isinstance(qkv_proj.scheme, CompressedTensorsW4A16Fp4)
+            assert qkv_proj.scheme.group_size == 16
+
+        llm.apply_model(check_model)
+        output = llm.generate_greedy("Hello my name is", max_tokens=20)
+        print(output)
+        assert output

vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py

@@ -23,9 +23,10 @@ from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tenso
     CompressedTensorsMoEMethod)
 from vllm.model_executor.layers.quantization.compressed_tensors.schemes import (
     W4A16SPARSE24_SUPPORTED_BITS, WNA16_SUPPORTED_BITS, CompressedTensors24,
-    CompressedTensorsScheme, CompressedTensorsW4A16Sparse24,
-    CompressedTensorsW8A8Fp8, CompressedTensorsW8A8Int8,
-    CompressedTensorsW8A16Fp8, CompressedTensorsWNA16)
+    CompressedTensorsScheme, CompressedTensorsW4A16Fp4,
+    CompressedTensorsW4A16Sparse24, CompressedTensorsW8A8Fp8,
+    CompressedTensorsW8A8Int8, CompressedTensorsW8A16Fp8,
+    CompressedTensorsWNA16)
 from vllm.model_executor.layers.quantization.compressed_tensors.utils import (
     find_matched_target, is_activation_quantization_format,
     should_ignore_layer)
@@ -216,6 +217,21 @@ class CompressedTensorsConfig(QuantizationConfig):
         else:
             return False
 
+    def _is_fp4a16_nvfp4(self, weight_quant: BaseModel,
+                         input_quant: BaseModel):
+
+        is_weight_only = weight_quant is not None and input_quant is None
+        is_group_quant = (
+            weight_quant.strategy == QuantizationStrategy.GROUP.value)
+        is_symmetric = weight_quant.symmetric
+
+        is_group_size_16 = weight_quant.group_size == 16
+        is_float_type = weight_quant.type == QuantizationType.FLOAT
+        is_4_bits = weight_quant.num_bits == 4
+
+        return (is_weight_only and is_group_quant and is_float_type
+                and is_4_bits and is_group_size_16 and is_symmetric)
+
     def _is_static_tensor_w8a8(self, weight_quant: BaseModel,
                                input_quant: BaseModel) -> bool:
         is_8_bits = weight_quant.num_bits == input_quant.num_bits == 8
@@ -315,6 +331,9 @@ class CompressedTensorsConfig(QuantizationConfig):
             input_quant: BaseModel) -> "CompressedTensorsScheme":
 
         # Detect If Mixed Precision
+        if self._is_fp4a16_nvfp4(weight_quant, input_quant):
+            return CompressedTensorsW4A16Fp4()
+
         if self._is_wNa16_group_channel(weight_quant, input_quant):
             if (self.quant_format == CompressionFormat.marlin_24.value
                     and weight_quant.num_bits in W4A16SPARSE24_SUPPORTED_BITS):

vllm/model_executor/layers/quantization/compressed_tensors/schemes/__init__.py

@@ -3,6 +3,7 @@
 from .compressed_tensors_scheme import CompressedTensorsScheme
 from .compressed_tensors_w4a16_24 import (W4A16SPARSE24_SUPPORTED_BITS,
                                           CompressedTensorsW4A16Sparse24)
+from .compressed_tensors_w4a16_nvfp4 import CompressedTensorsW4A16Fp4
 from .compressed_tensors_w8a8_fp8 import CompressedTensorsW8A8Fp8
 from .compressed_tensors_w8a8_int8 import CompressedTensorsW8A8Int8
 from .compressed_tensors_w8a16_fp8 import CompressedTensorsW8A16Fp8
@@ -16,5 +17,5 @@ __all__ = [
     "CompressedTensorsW8A16Fp8", "CompressedTensorsW4A16Sparse24",
     "CompressedTensorsW8A8Int8", "CompressedTensorsW8A8Fp8",
     "WNA16_SUPPORTED_BITS", "W4A16SPARSE24_SUPPORTED_BITS",
-    "CompressedTensors24"
+    "CompressedTensors24", "CompressedTensorsW4A16Fp4"
 ]

vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a16_nvfp4.py

@@ -0,0 +1,107 @@
+# SPDX-License-Identifier: Apache-2.0
+from typing import Callable, List, Optional
+
+import torch
+import torch.nn.functional as F
+from torch.nn.parameter import Parameter
+
+from vllm.model_executor.layers.quantization.compressed_tensors.schemes import (
+    CompressedTensorsScheme)
+from vllm.model_executor.layers.quantization.utils.nvfp4_emulation_utils import (  # noqa: E501
+    dequantize_to_dtype)
+from vllm.model_executor.parameter import (GroupQuantScaleParameter,
+                                           ModelWeightParameter,
+                                           PerTensorScaleParameter)
+
+__all__ = ["CompressedTensorsW4A16Fp4"]
+
+
+class CompressedTensorsW4A16Fp4(CompressedTensorsScheme):
+
+    def __init__(self):
+        self.group_size = 16
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        # dont restrict as emulations
+        return 80
+
+    def create_weights(self, layer: torch.nn.Module,
+                       output_partition_sizes: List[int],
+                       input_size_per_partition: int,
+                       params_dtype: torch.dtype, weight_loader: Callable,
+                       **kwargs):
+
+        # Weight
+        weight = ModelWeightParameter(data=torch.empty(
+            sum(output_partition_sizes),
+            input_size_per_partition // 2,
+            dtype=torch.uint8),
+                                      input_dim=1,
+                                      output_dim=0,
+                                      weight_loader=weight_loader)
+        layer.register_parameter("weight_packed", weight)
+
+        # Global Weight Scale
+        weight_global_scale = PerTensorScaleParameter(
+            data=torch.empty(len(output_partition_sizes), dtype=torch.float32),
+            weight_loader=weight_loader)
+        layer.register_parameter("weight_global_scale", weight_global_scale)
+
+        # Per Group Weight Scale
+        weight_scale = GroupQuantScaleParameter(data=torch.empty(
+            sum(output_partition_sizes),
+            input_size_per_partition // self.group_size,
+            dtype=torch.float8_e4m3fn,
+        ),
+                                                input_dim=1,
+                                                output_dim=0,
+                                                weight_loader=weight_loader)
+
+        layer.register_parameter("weight_scale", weight_scale)
+
+    def swizzle_blockscale(self, scale: torch.tensor):
+        assert (scale.dtype == torch.float8_e4m3fn)
+        # Pad and blockwise interleave weight_scale
+        scale_ndim = scale.ndim
+        if scale.ndim == 2:
+            scale = scale.unsqueeze(0)
+        assert scale.ndim == 3
+        B, M, K = scale.shape
+        round_up_multiple = lambda x, m: (x + m - 1) // m * m
+        M_padded = round_up_multiple(M, 128)
+        K_padded = round_up_multiple(K, 4)
+        padded_scale = torch.zeros((B, M_padded, K_padded), dtype=scale.dtype)
+        padded_scale[:B, :M, :K] = scale
+        batches, rows, cols = padded_scale.shape
+        assert rows % 128 == 0
+        assert cols % 4 == 0
+        padded_scale = padded_scale.reshape(batches, rows // 128, 4, 32,
+                                            cols // 4, 4)
+        swizzled_scale = padded_scale.permute((0, 1, 4, 3, 2, 5))
+        swizzled_scale = swizzled_scale.contiguous().cuda()
+        return (swizzled_scale.reshape(M, K)
+                if scale_ndim == 2 else swizzled_scale.reshape(B, M, K))
+
+    def process_weights_after_loading(self, layer) -> None:
+        layer.weight_global_scale = Parameter(
+            layer.weight_global_scale.max().to(torch.float32),
+            requires_grad=False)
+        # Note: a post weight loading step but not required for the emulation
+        swizzled_weight_scale = self.swizzle_blockscale(layer.weight_scale)
+        layer.weight_scale_swizzled = Parameter(swizzled_weight_scale,
+                                                requires_grad=False)
+
+    def apply_weights(self,
+                      layer: torch.nn.Module,
+                      x: torch.Tensor,
+                      bias: Optional[torch.Tensor] = None) -> torch.Tensor:
+
+        w_fp4 = layer.weight_packed.data
+        w_global_scale = layer.weight_global_scale
+        w_blockscale = layer.weight_scale_swizzled.data
+        w_dq = dequantize_to_dtype(w_fp4, w_blockscale, w_global_scale,
+                                   x.dtype, x.device, self.group_size)
+        out = F.linear(x, w_dq)
+        del w_dq, w_fp4, w_global_scale, w_blockscale
+        return out

vllm/model_executor/layers/quantization/utils/nvfp4_emulation_utils.py

@@ -0,0 +1,61 @@
+# SPDX-License-Identifier: Apache-2.0
+import torch
+
+__all__ = [
+    "break_fp4_bytes",
+    "dequantize_to_dtype",
+]
+
+kE2M1ToFloat = torch.tensor([0., 0.5, 1., 1.5, 2., 3., 4., 6.],
+                            dtype=torch.float32)
+
+
+def break_fp4_bytes(a, dtype):
+    assert a.dtype == torch.uint8
+    m, n = a.shape
+    # Vectorized nibble processing
+    a_flat = a.flatten()
+    high = (a_flat & 0xF0) >> 4  # Upper nibbles
+    low = a_flat & 0x0F  # Lower nibbles
+    # Combine nibbles for batch processing
+    combined = torch.stack((low, high), dim=1).flatten()
+    # Vectorized sign and magnitude extraction
+    signs = (combined & 0x08).to(torch.bool)  # Sign bits
+    abs_vals = (combined & 0x07).to(torch.long)
+    # Device-aware lookup and sign application
+    kE2M1 = kE2M1ToFloat.to(device=a.device)
+    values = kE2M1[abs_vals] * torch.where(signs, -1.0, 1.0)
+    # Reshape to final form
+    return values.reshape(m, n * 2).to(dtype=dtype)
+
+
+def convert_swizzled_to_linear(a_sf_swizzled: torch.Tensor, m, k, block_size):
+    m_tiles = (m + 128 - 1) // 128
+    f = block_size * 4
+    k_tiles = (k + f - 1) // f
+    tmp = torch.reshape(a_sf_swizzled, (1, m_tiles, k_tiles, 32, 4, 4))
+    tmp = torch.permute(tmp, (0, 1, 4, 3, 2, 5))
+    out = tmp.reshape(m_tiles * 128, k_tiles * f // block_size)
+    return out[0:m, 0:k]
+
+
+def dequantize_to_dtype(tensor_fp4,
+                        tensor_sf,
+                        global_scale,
+                        dtype,
+                        device,
+                        block_size=16):
+    """Dequantize the fp4 tensor back to high precision."""
+    # Two fp4 values are packed into one uint8.
+    assert tensor_fp4.dtype == torch.uint8
+    m, packed_k = tensor_fp4.shape
+    k = packed_k * 2
+    tensor_f32 = break_fp4_bytes(tensor_fp4, torch.float32)
+    tensor_f32 = tensor_f32.reshape(m, k // block_size, block_size)
+    tensor_sf = tensor_sf.view(torch.float8_e4m3fn)
+    tensor_sf = convert_swizzled_to_linear(tensor_sf, m, k, block_size)
+    tensor_sf_dtype = tensor_sf.to(torch.float32) / global_scale
+
+    # scale the tensor
+    out = (tensor_f32 * tensor_sf_dtype.unsqueeze(-1)).reshape(m, k)
+    return out.to(dtype)