vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Massive speed improvement thanks to Cebtenzzre

Browse source
This commit is contained in:
KerfuffleV2 2023-08-24 12:13:00 -06:00
parent 3efcbb8f59
commit 8ee186c0aa
1 changed files with 9 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

16
convert.py
View file

@ -764,13 +764,15 @@ def quantize_array_q8_0(arr):
assert arr.dtype == np.float32, f'Bad array type {arr.dtype}' assert arr.dtype == np.float32, f'Bad array type {arr.dtype}'
n_blocks = arr.size // QK8_0 n_blocks = arr.size // QK8_0
blocks = arr.reshape((n_blocks, QK8_0)) blocks = arr.reshape((n_blocks, QK8_0))
return np.fromiter(map(quantize_block_q8_0, blocks), count = n_blocks, dtype = BLOCK_Q8_0) return np.fromiter(quantize_blocks_q8_0(blocks), count = n_blocks, dtype = BLOCK_Q8_0)
def quantize_block_q8_0(blk, zero = np.float32(0), one = np.float32(1), onetwentyseven = np.float32(127), zero_chunk = (np.int8(0),) * QK8_0): # Much faster implementation of block quantization contributed by @Cebtenzzre
d = abs(blk).max() / onetwentyseven def quantize_blocks_q8_0(blocks):
if d == zero: d = abs(blocks).max(axis = 1) / np.float32(127)
return (np.float16(d), zero_chunk) with np.errstate(divide = 'ignore'):
return (np.float16(d), (blk * (one / d)).round()) qs = (blocks / d[:, None]).round()
qs[d == 0] = 0
yield from zip(np.float16(d), qs)
class OutputFile: class OutputFile:
@ -892,7 +894,7 @@ class OutputFile:
elapsed = time.time() - start elapsed = time.time() - start
size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape) size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape)
padi = len(str(len(model))) padi = len(str(len(model)))
print(f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:6} | T+{int(elapsed):4}") print(f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}")
of.gguf.write_tensor_data(ndarray) of.gguf.write_tensor_data(ndarray)
of.close() of.close()