Skip to content
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

Added GGML Type 3 dequantization capability in transformers #32062

Closed
wants to merge 4 commits into from
Closed

Added GGML Type 3 dequantization capability in transformers #32062

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
69739f9
Added dequantize_q4_1 function to ggml
kskd1804 Jul 18, 2024
b0e94fc
Added ggml type for q4_1
kskd1804 Jul 18, 2024
31ec19c
Added calls to dequantize_q4_1 function
kskd1804 Jul 18, 2024
3def5bc
Merge branch 'huggingface:main' into dequantize_q4_1_fn
kskd1804 Jul 18, 2024
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
34 changes: 34 additions & 0 deletions src/transformers/integrations/ggml.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,6 +37,7 @@
GGML_TYPES = {
"F32": 0,
"Q4_0": 2,
"Q4_1": 3,
"Q8_0": 8,
"Q2_K": 10,
"Q3_K": 11,
Expand All @@ -52,6 +53,7 @@
"Q4_K": 144,
# Q4_0 uses a blocksize of 32 but the 4-bit tensors are packed into 8-bit tensors + 2 bytes for the scales
"Q4_0": 2 + 16,
"Q4_1": 2 + 2 + 16,
"Q6_K": 210,
# See: https://github.com/99991/pygguf/commit/a417edbfc029a1bc270f984a694f9128c5afa8b9
"Q2_K": 256 // 16 + 256 // 4 + 2 + 2,
Expand Down Expand Up @@ -273,6 +275,36 @@ def dequantize_q4_0(data):
return (scales * quants).astype(np.float32)


def dequantize_q4_1(data):
# C implementation
# https://github.com/ggerganov/ggml/blob/fca1caafea7de9fbd7efc733b9818f9cf2da3050/src/ggml-quants.c#L1106
# C struct definition
# https://github.com/ggerganov/ggml/blob/fca1caafea7de9fbd7efc733b9818f9cf2da3050/src/ggml-quants.h#L18
block_size = GGML_BLOCK_SIZES["Q4_1"]
num_blocks = len(data) // block_size

data_f16 = np.frombuffer(data, dtype=np.float16).reshape(num_blocks, block_size // 2)
data_u8 = np.frombuffer(data, dtype=np.uint8).reshape(num_blocks, block_size)

# The scales are stored on the first 2 bytes
scales = data_f16[:, 0].reshape(num_blocks, 1).astype(np.float32)
# scales = np.nan_to_num(scales)

# The mins are stored on the second 2 bytes
mins = data_f16[:, 1].reshape(num_blocks, 1).astype(np.float32)

# the rest of the bytes corresponds to the quants - we discard the first four bytes
quants = data_u8[:, 4:]

ql = (quants[:, :] & 0xF).astype(np.int8)
qr = (quants[:, :] >> 4).astype(np.int8)

# Use hstack
quants = np.hstack([ql, qr])

return ((scales * quants) + mins).astype(np.float32)


def dequantize_q6_k(data):
# C implementation
# https://github.com/ggerganov/ggml/blob/fca1caafea7de9fbd7efc733b9818f9cf2da3050/src/ggml-quants.c#L2275
Expand Down Expand Up @@ -493,6 +525,8 @@ def load_dequant_gguf_tensor(shape, ggml_type, data):
values = dequantize_q8_0(data)
elif ggml_type == GGML_TYPES["Q4_0"]:
values = dequantize_q4_0(data)
elif ggml_type == GGML_TYPES["Q4_1"]:
values = dequantize_q4_1(data)
elif ggml_type == GGML_TYPES["Q4_K"]:
values = dequantize_q4_k(data)
elif ggml_type == GGML_TYPES["Q6_K"]:
Expand Down