Make loras work on nvfp4 models. (#11837)

The initial applying is a bit slow but will probably be sped up in the
future.

comfy/float.py

@@ -65,3 +65,116 @@ def stochastic_rounding(value, dtype, seed=0):
         return output
 
     return value.to(dtype=dtype)
+
+
+# TODO: improve this?
+def stochastic_float_to_fp4_e2m1(x, generator):
+    sign = torch.signbit(x).to(torch.uint8)
+    x_abs = x.abs()
+
+    exp = torch.floor(torch.log2(x_abs) + 1.0).clamp(0, 3)
+    x += (torch.rand(x.size(), dtype=x.dtype, layout=x.layout, device=x.device, generator=generator) - 0.5) * (2 ** (exp - 2.0)) * 1.25
+
+    x_abs = x.abs()
+    exp = torch.floor(torch.log2(x_abs) + 1.1925).clamp(0, 3)
+
+    mantissa = torch.where(
+        exp > 0,
+        (x_abs / (2.0 ** (exp - 1)) - 1.0) * 2.0,
+        (x_abs * 2.0)
+    ).round().to(torch.uint8)
+
+    fp4 = (sign << 3) | (exp.to(torch.uint8) << 1) | mantissa
+
+    fp4_flat = fp4.view(-1)
+    packed = (fp4_flat[0::2] << 4) | fp4_flat[1::2]
+    return packed.reshape(list(x.shape)[:-1] + [-1])
+
+
+def to_blocked(input_matrix, flatten: bool = True) -> torch.Tensor:
+    """
+    Rearrange a large matrix by breaking it into blocks and applying the rearrangement pattern.
+    See:
+        https://docs.nvidia.com/cuda/cublas/index.html#d-block-scaling-factors-layout
+
+    Args:
+        input_matrix: Input tensor of shape (H, W)
+    Returns:
+        Rearranged tensor of shape (32*ceil_div(H,128), 16*ceil_div(W,4))
+    """
+
+    def ceil_div(a, b):
+        return (a + b - 1) // b
+
+    rows, cols = input_matrix.shape
+    n_row_blocks = ceil_div(rows, 128)
+    n_col_blocks = ceil_div(cols, 4)
+
+    # Calculate the padded shape
+    padded_rows = n_row_blocks * 128
+    padded_cols = n_col_blocks * 4
+
+    padded = input_matrix
+    if (rows, cols) != (padded_rows, padded_cols):
+        padded = torch.zeros(
+            (padded_rows, padded_cols),
+            device=input_matrix.device,
+            dtype=input_matrix.dtype,
+        )
+        padded[:rows, :cols] = input_matrix
+
+    # Rearrange the blocks
+    blocks = padded.view(n_row_blocks, 128, n_col_blocks, 4).permute(0, 2, 1, 3)
+    rearranged = blocks.reshape(-1, 4, 32, 4).transpose(1, 2).reshape(-1, 32, 16)
+    if flatten:
+        return rearranged.flatten()
+
+    return rearranged.reshape(padded_rows, padded_cols)
+
+
+def stochastic_round_quantize_nvfp4(x, per_tensor_scale, pad_16x, seed=0):
+    F4_E2M1_MAX = 6.0
+    F8_E4M3_MAX = 448.0
+
+    def roundup(x: int, multiple: int) -> int:
+        """Round up x to the nearest multiple."""
+        return ((x + multiple - 1) // multiple) * multiple
+
+    orig_shape = x.shape
+
+    # Handle padding
+    if pad_16x:
+        rows, cols = x.shape
+        padded_rows = roundup(rows, 16)
+        padded_cols = roundup(cols, 16)
+        if padded_rows != rows or padded_cols != cols:
+            x = torch.nn.functional.pad(x, (0, padded_cols - cols, 0, padded_rows - rows))
+            # Note: We update orig_shape because the output tensor logic below assumes x.shape matches
+            # what we want to produce. If we pad here, we want the padded output.
+            orig_shape = x.shape
+
+    block_size = 16
+
+    x = x.reshape(orig_shape[0], -1, block_size)
+    max_abs = torch.amax(torch.abs(x), dim=-1)
+    block_scale = max_abs / F4_E2M1_MAX
+    scaled_block_scales = block_scale / per_tensor_scale.to(block_scale.dtype)
+    scaled_block_scales_fp8 = torch.clamp(scaled_block_scales, max=F8_E4M3_MAX).to(torch.float8_e4m3fn)
+    total_scale = per_tensor_scale.to(x.dtype) * scaled_block_scales_fp8.to(x.dtype)
+
+    # Handle zero blocks (from padding): avoid 0/0 NaN
+    zero_scale_mask = (total_scale == 0)
+    total_scale_safe = torch.where(zero_scale_mask, torch.ones_like(total_scale), total_scale)
+
+    x = x / total_scale_safe.unsqueeze(-1)
+
+    generator = torch.Generator(device=x.device)
+    generator.manual_seed(seed)
+
+    x = torch.where(zero_scale_mask.unsqueeze(-1), torch.zeros_like(x), x)
+
+    x = x.view(orig_shape)
+    data_lp = stochastic_float_to_fp4_e2m1(x, generator=generator)
+
+    blocked_scales = to_blocked(scaled_block_scales_fp8, flatten=False)
+    return data_lp, blocked_scales

comfy/ops.py

@@ -699,7 +699,7 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
             def set_weight(self, weight, inplace_update=False, seed=None, return_weight=False, **kwargs):
                 if getattr(self, 'layout_type', None) is not None:
                     # dtype is now implicit in the layout class
-                    weight = QuantizedTensor.from_float(weight, self.layout_type, scale="recalculate", stochastic_rounding=seed, inplace_ops=True)
+                    weight = QuantizedTensor.from_float(weight, self.layout_type, scale="recalculate", stochastic_rounding=seed, inplace_ops=True).to(self.weight.dtype)
                 else:
                     weight = weight.to(self.weight.dtype)
                 if return_weight:

comfy/quant_ops.py

@@ -7,7 +7,7 @@ try:
         QuantizedTensor,
         QuantizedLayout,
         TensorCoreFP8Layout as _CKFp8Layout,
-        TensorCoreNVFP4Layout,  # Direct import, no wrapper needed
+        TensorCoreNVFP4Layout as _CKNvfp4Layout,
         register_layout_op,
         register_layout_class,
         get_layout_class,
@@ -34,7 +34,7 @@ except ImportError as e:
     class _CKFp8Layout:
         pass
 
-    class TensorCoreNVFP4Layout:
+    class _CKNvfp4Layout:
         pass
 
     def register_layout_class(name, cls):
@@ -84,6 +84,39 @@ class _TensorCoreFP8LayoutBase(_CKFp8Layout):
         return qdata, params
 
 
+class TensorCoreNVFP4Layout(_CKNvfp4Layout):
+    @classmethod
+    def quantize(cls, tensor, scale=None, stochastic_rounding=0, inplace_ops=False):
+        if tensor.dim() != 2:
+            raise ValueError(f"NVFP4 requires 2D tensor, got {tensor.dim()}D")
+
+        orig_dtype = tensor.dtype
+        orig_shape = tuple(tensor.shape)
+
+        if scale is None or (isinstance(scale, str) and scale == "recalculate"):
+            scale = torch.amax(tensor.abs()) / (ck.float_utils.F8_E4M3_MAX * ck.float_utils.F4_E2M1_MAX)
+
+        if not isinstance(scale, torch.Tensor):
+            scale = torch.tensor(scale)
+        scale = scale.to(device=tensor.device, dtype=torch.float32)
+
+        padded_shape = cls.get_padded_shape(orig_shape)
+        needs_padding = padded_shape != orig_shape
+
+        if stochastic_rounding > 0:
+            qdata, block_scale = comfy.float.stochastic_round_quantize_nvfp4(tensor, scale, pad_16x=needs_padding, seed=stochastic_rounding)
+        else:
+            qdata, block_scale = ck.quantize_nvfp4(tensor, scale, pad_16x=needs_padding)
+
+        params = cls.Params(
+            scale=scale,
+            orig_dtype=orig_dtype,
+            orig_shape=orig_shape,
+            block_scale=block_scale,
+        )
+        return qdata, params
+
+
 class TensorCoreFP8E4M3Layout(_TensorCoreFP8LayoutBase):
     FP8_DTYPE = torch.float8_e4m3fn
 

requirements.txt

@@ -21,7 +21,7 @@ psutil
 alembic
 SQLAlchemy
 av>=14.2.0
-comfy-kitchen>=0.2.5
+comfy-kitchen>=0.2.6
 
 #non essential dependencies:
 kornia>=0.7.1