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

LLM: Add new qtype woq_int4 to support gemm int4 temporary. #12706

Merged
merged 1 commit into from
Jan 15, 2025
Merged

LLM: Add new qtype woq_int4 to support gemm int4 temporary. #12706

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
1 change: 1 addition & 0 deletions python/llm/src/ipex_llm/ggml/quantize.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -53,6 +53,7 @@
"sym_int4_rtn": 31,
"sym_int8_rtn": 32,
"asym_int4_rtn": 33,
"woq_int4": 34,
}

# mixed precison from llama.cpp
Expand Down
5 changes: 3 additions & 2 deletions python/llm/src/ipex_llm/transformers/low_bit_linear.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -84,6 +84,7 @@
SYM_INT4_RTN = ggml_tensor_qtype["sym_int4_rtn"]
SYM_INT8_RTN = ggml_tensor_qtype["sym_int8_rtn"]
ASYM_INT4_RTN = ggml_tensor_qtype["asym_int4_rtn"]
WOQ_INT4 = ggml_tensor_qtype["woq_int4"]
RTN_DTYPE = {
SYM_INT4_RTN: torch.uint8,
ASYM_INT4_RTN: torch.uint8,
Expand Down Expand Up @@ -187,7 +188,7 @@ def ggml_q_format_convet_cpu2xpu(tensor: torch.Tensor, num_elem: int, qtype: int
src = ctypes.c_void_p(tensor.data.data_ptr())

if qtype in [SYM_INT4, ASYM_INT4, SYM_INT8, NF4, NF3, FP4, FP6, FP8E4, FP8E5,
Q4_K, Q6_K, FP6_K]:
Q4_K, Q6_K, FP6_K, WOQ_INT4]:
dst_tensor = torch.empty_like(tensor)
elif qtype == ggml_tensor_qtype["sym_int5"]:
QK = ggml.ggml_qk_size(qtype)
Expand All @@ -213,7 +214,7 @@ def ggml_q_format_convet_xpu2cpu(tensor: torch.Tensor, num_elem: int, qtype: int
src = ctypes.c_void_p(tensor.data.data_ptr())

if qtype in [SYM_INT4, ASYM_INT4, SYM_INT8, NF4, NF3, FP4, FP6, FP8E4, FP8E5,
Q4_K, Q6_K, FP6_K]:
Q4_K, Q6_K, FP6_K, WOQ_INT4]:
dst_tensor = torch.empty_like(tensor)
elif qtype == ggml_tensor_qtype["sym_int5"]:
QK = ggml.ggml_qk_size(ggml_tensor_qtype["asym_int5"])
Expand Down
8 changes: 4 additions & 4 deletions python/llm/src/ipex_llm/transformers/models/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,7 +21,7 @@
from ipex_llm.ggml.quantize import ggml_tensor_qtype
from ipex_llm.transformers.utils import get_xpu_device_name
from ipex_llm.transformers.low_bit_linear import SYM_INT4, SYM_INT8, FP8E5, IQ2_XXS, FP4, FP8E4,\
FP6, ASYM_INT4
FP6, ASYM_INT4, WOQ_INT4

FP8_KV_ALLOC_LENGTH = 512
KV_CACHE_ALLOC_BLOCK_LENGTH = int(os.environ.get("KV_CACHE_ALLOC_BLOCK_LENGTH", 256))
Expand All @@ -33,7 +33,7 @@

def decoding_fast_path_qtype_check(proj):
qtype = getattr(proj, "qtype", None)
return qtype in [SYM_INT4, FP8E5, FP4]
return qtype in [SYM_INT4, FP8E5, FP4, WOQ_INT4]


def init_kv_cache(batch_size, num_heads, head_dim, current_length, max_length, dtype, device):
Expand Down Expand Up @@ -248,7 +248,7 @@ def mlp_fusion_check(x, qtype, training):
return False
if x.device.type != 'xpu':
return False
if qtype not in [SYM_INT4, FP8E5, FP4, IQ2_XXS, FP6]:
if qtype not in [SYM_INT4, FP8E5, FP4, IQ2_XXS, FP6, WOQ_INT4]:
return False
if training or x.requires_grad:
return False
Expand All @@ -263,7 +263,7 @@ def use_xmx(x: torch.Tensor, qtype: int):
device = get_xpu_device_name(x.device)
return (
device in ["arc", "pvc"]
and qtype in [SYM_INT4, SYM_INT8, FP8E4, FP8E5]
and qtype in [SYM_INT4, SYM_INT8, FP8E4, FP8E5, WOQ_INT4]
and (
(device == "pvc" and 1 < x.size(0) <= 16)
or
Expand Down
Loading