ayo/ipex-llm
1
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

Add NF3 (#9008)

Browse source 
* add nf3

* grammar
This commit is contained in:
Kai Huang 2023-09-20 20:03:07 +08:00 committed by GitHub
parent 6981745fe4
commit 156af15d1e
2 changed files with 7 additions and 5 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

3
python/llm/src/bigdl/llm/ggml/quantize.py
View file

@ -30,7 +30,8 @@ ggml_tensor_qtype = {"sym_int4": 2, # q4_0 in ggml
"sym_int5": 6, # q5_0 in ggml "sym_int5": 6, # q5_0 in ggml
"asym_int5": 7, # q5_1 in ggml "asym_int5": 7, # q5_1 in ggml
"sym_int8": 8, # q8_0 in ggml "sym_int8": 8, # q8_0 in ggml
"nf4": 10} "nf4": 10,
"nf3": 11}
_llama_quantize_type = {"q4_0": 2, _llama_quantize_type = {"q4_0": 2,
"q4_1": 3, "q4_1": 3,

9
python/llm/src/bigdl/llm/transformers/low_bit_linear.py
View file

@ -62,6 +62,7 @@ TORCH_LINEAR_THRESHOLD = 96
SYM_INT4 = ggml_tensor_qtype["sym_int4"] SYM_INT4 = ggml_tensor_qtype["sym_int4"]
SYM_INT8 = ggml_tensor_qtype["sym_int8"] SYM_INT8 = ggml_tensor_qtype["sym_int8"]
NF4 = ggml_tensor_qtype["nf4"] NF4 = ggml_tensor_qtype["nf4"]
NF3 = ggml_tensor_qtype["nf3"]
def ggml_convert_qtype(tensor: torch.Tensor, qtype: int, def ggml_convert_qtype(tensor: torch.Tensor, qtype: int,
@ -101,7 +102,7 @@ def ggml_q_format_convet_cpu2xpu(tensor: torch.Tensor, num_elem: int, qtype: int
src = ctypes.c_void_p(tensor.data.data_ptr()) src = ctypes.c_void_p(tensor.data.data_ptr())
if qtype in [SYM_INT4, SYM_INT8, NF4]: if qtype in [SYM_INT4, SYM_INT8, NF4, NF3]:
dst_tensor = torch.empty_like(tensor) dst_tensor = torch.empty_like(tensor)
elif qtype == ggml_tensor_qtype["sym_int5"]: elif qtype == ggml_tensor_qtype["sym_int5"]:
QK = ggml.ggml_qk_size(qtype) QK = ggml.ggml_qk_size(qtype)
@ -126,7 +127,7 @@ def ggml_q_format_convet_xpu2cpu(tensor: torch.Tensor, num_elem: int, qtype: int
src = ctypes.c_void_p(tensor.data.data_ptr()) src = ctypes.c_void_p(tensor.data.data_ptr())
if qtype in [SYM_INT4, SYM_INT8, NF4]: if qtype in [SYM_INT4, SYM_INT8, NF4, NF3]:
dst_tensor = torch.empty_like(tensor) dst_tensor = torch.empty_like(tensor)
elif qtype == ggml_tensor_qtype["sym_int5"]: elif qtype == ggml_tensor_qtype["sym_int5"]:
QK = ggml.ggml_qk_size(ggml_tensor_qtype["asym_int5"]) QK = ggml.ggml_qk_size(ggml_tensor_qtype["asym_int5"])
@ -363,8 +364,8 @@ class LowBitLinear(nn.Linear):
else: else:
# CPU logic # CPU logic
# todo may need to set a different number on different platforms # todo may need to set a different number on different platforms
invalidInputError(self.qtype != ggml_tensor_qtype["nf4"], invalidInputError(self.qtype != NF3 and self.qtype != NF4,
"NF4 quantization is currently not supported on CPU") "NF3 and NF4 quantization are currently not supported on CPU")
if IS_SERVER and (not IS_SPR) and \ if IS_SERVER and (not IS_SPR) and \
self.qtype == SYM_INT4 and x_2d.shape[0] >= TORCH_LINEAR_THRESHOLD: self.qtype == SYM_INT4 and x_2d.shape[0] >= TORCH_LINEAR_THRESHOLD:
x0_fp32 = ggml_int4_convert_fp32(x0, self.weight_shape, self.weight_length) x0_fp32 = ggml_int4_convert_fp32(x0, self.weight_shape, self.weight_length)