Skip to content

gguf new quant type support (with demo) #12076

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Sep 9, 2025
Merged

gguf new quant type support (with demo) #12076

Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
15adb60
Update utils.py
calcuis Aug 3, 2025
bc6bd9a
Merge branch 'huggingface:main' into main
calcuis Sep 9, 2025
c80f89d
Apply style fixes
github-actions[bot] Sep 9, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
48 changes: 48 additions & 0 deletions src/diffusers/quantizers/gguf/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -428,9 +428,57 @@ def dequantize_blocks_Q2_K(blocks, block_size, type_size, dtype=None):
def dequantize_blocks_BF16(blocks, block_size, type_size, dtype=None):
return (blocks.view(torch.int16).to(torch.int32) << 16).view(torch.float32)

# this part from calcuis (gguf.org)
# more info: https://github.com/calcuis/gguf-connector/blob/main/src/gguf_connector/quant2c.py

def dequantize_blocks_IQ4_NL(blocks, block_size, type_size, dtype=None):
kvalues = torch.tensor(
[-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113],
dtype=torch.float32, device=blocks.device
)
n_blocks = blocks.shape[0]
d, qs = split_block_dims(blocks, 2)
d = d.view(torch.float16).to(dtype)
qs = qs.reshape((n_blocks, -1, 1, block_size // 2)) >> torch.tensor(
[0, 4], device=blocks.device, dtype=torch.uint8
).reshape((1, 1, 2, 1))
qs = (qs & 15).reshape((n_blocks, -1)).to(torch.int64)
kvalues = kvalues.view(1, 1, 16)
qs = qs.unsqueeze(-1)
qs = torch.gather(kvalues.expand(qs.shape[0], qs.shape[1], 16), 2, qs)
qs = qs.squeeze(-1).to(dtype)
return d * qs

def dequantize_blocks_IQ4_XS(blocks, block_size, type_size, dtype=None):
kvalues = torch.tensor(
[-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113],
dtype=torch.float32, device=blocks.device
)
n_blocks = blocks.shape[0]
d, scales_h, scales_l, qs = split_block_dims(blocks, 2, 2, QK_K // 64)
d = d.view(torch.float16).to(dtype)
scales_h = scales_h.view(torch.int16)
scales_l = scales_l.reshape((n_blocks, -1, 1)) >> torch.tensor(
[0, 4], device=blocks.device, dtype=torch.uint8).reshape((1, 1, 2))
scales_h = scales_h.reshape((n_blocks, 1, -1)) >> torch.tensor(
[2 * i for i in range(QK_K // 32)], device=blocks.device, dtype=torch.uint8).reshape((1, -1, 1))
scales_l = scales_l.reshape((n_blocks, -1)) & 0x0F
scales_h = scales_h.reshape((n_blocks, -1)) & 0x03
scales = (scales_l | (scales_h << 4)) - 32
dl = (d * scales.to(dtype)).reshape((n_blocks, -1, 1))
shifts_q = torch.tensor([0, 4], device=blocks.device, dtype=torch.uint8).reshape(1, 1, 2, 1)
qs = qs.reshape((n_blocks, -1, 1, 16)) >> shifts_q
qs = (qs & 15).reshape((n_blocks, -1, 32)).to(torch.int64)
kvalues = kvalues.view(1, 1, 1, 16)
qs = qs.unsqueeze(-1)
qs = torch.gather(kvalues.expand(qs.shape[0], qs.shape[1], qs.shape[2], 16), 3, qs)
qs = qs.squeeze(-1).to(dtype)
return (dl * qs).reshape(n_blocks, -1)

GGML_QUANT_SIZES = gguf.GGML_QUANT_SIZES
dequantize_functions = {
gguf.GGMLQuantizationType.IQ4_NL: dequantize_blocks_IQ4_NL,
gguf.GGMLQuantizationType.IQ4_XS: dequantize_blocks_IQ4_XS,
gguf.GGMLQuantizationType.BF16: dequantize_blocks_BF16,
gguf.GGMLQuantizationType.Q8_0: dequantize_blocks_Q8_0,
gguf.GGMLQuantizationType.Q5_1: dequantize_blocks_Q5_1,
Expand Down
Loading