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

[NPU] support asym_int4 for baichuan #12576

Draft
wants to merge 1 commit into
base: main
Choose a base branch
from
Draft

[NPU] support asym_int4 for baichuan #12576

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
48 changes: 37 additions & 11 deletions python/llm/src/ipex_llm/transformers/npu_models/baichuan_mp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -80,7 +80,8 @@ def __init__(
intermediate_size,
n_splits_linear: int = 1,
n_splits_down_proj: int = 1,
group_size: int = 0
group_size: int = 0,
asym: bool = False,
):
super().__init__(max_seq_len=max_seq_len,
transpose_value=transpose_value,
Expand All @@ -89,7 +90,8 @@ def __init__(
device=device,
n_splits_linear=n_splits_linear,
n_splits_down_proj=n_splits_down_proj,
group_size=group_size)
group_size=group_size,
asym=asym)
self.max_seq_len = max_seq_len
self.intermediate_size = intermediate_size
self.dtype = dtype
Expand Down Expand Up @@ -368,16 +370,19 @@ def __init__(
do_print: bool = False,
n_splits_linear: int = 1,
n_splits_down_proj: int = 1,
group_size: int = 0
group_size: int = 0,
asym: bool = False,
):
super().__init__()

self.do_print = do_print

op_parameters = []
for w in parameters:
if isinstance(w, tuple): # from QuantizedLinear
if isinstance(w, tuple) and not asym: # from QuantizedLinear
op_parameters.append((w[0].numpy(), w[1].numpy()))
elif isinstance(w, tuple) and asym: # from QuantizedLinear
op_parameters.append((w[0].numpy(), w[1].numpy(), w[2].numpy()))
elif w.dtype in [torch.int8, torch.uint8]: # QuantizedLinear weight
op_parameters.append(w.numpy())
elif isinstance(w, np.ndarray): # scale
Expand Down Expand Up @@ -430,7 +435,8 @@ def __init__(
dtype=np_dtype,
n_splits_linear=n_splits_linear,
n_splits_down_proj=n_splits_down_proj,
group_size=group_size
group_size=group_size,
asym=asym,
)
self.backend_decoders.append(decoder)

Expand Down Expand Up @@ -506,7 +512,8 @@ def __init__(
transpose_value: bool = False,
n_splits_linear: int = 1,
n_splits_down_proj: int = 1,
group_size: int = 0
group_size: int = 0,
asym: bool = False,
):
super().__init__()
self.op_parameters = parameters
Expand Down Expand Up @@ -537,7 +544,8 @@ def __init__(
dtype=np_dtype,
n_splits_linear=n_splits_linear,
n_splits_down_proj=n_splits_down_proj,
group_size=group_size
group_size=group_size,
asym=asym
)
self.layer_norm_0 = layer_norm_0
self.layer_norm_1 = layer_norm_1
Expand Down Expand Up @@ -620,6 +628,7 @@ def run_decode(
layer_indexs = range(layer_start, layer_end)
n_splits_linear = len(model.model.layers[0].mlp.gate_proj_dq_list)
n_splits_down_proj = len(model.model.layers[0].mlp.down_proj_dq_list)
asym = getattr(model.config, "asym", False)
for layer_idx in layer_indexs:
curr_layer = model.model.layers[layer_idx]
attn_layer = curr_layer.self_attn
Expand All @@ -631,10 +640,17 @@ def run_decode(
mlp_layer.down_proj_dq_list]:
l_weights = []
scales = []
zeros = []
for l in layer_list:
l_weights.append(l.weight)
scales.append(l.scale)
weights.append((torch.stack(l_weights, axis=0), torch.stack(scales, axis=0)))
if l.zero is not None:
zeros.append(l.zero)
if len(zeros):
weights.append((torch.stack(l_weights, axis=0), torch.stack(scales, axis=0),
torch.stack(zeros, axis=0)))
else:
weights.append((torch.stack(l_weights, axis=0), torch.stack(scales, axis=0)))

cached_cos = curr_layer.self_attn.rotary_emb.cos_cached.to(torch.float16)
cached_sin = curr_layer.self_attn.rotary_emb.sin_cached.to(torch.float16)
Expand Down Expand Up @@ -663,7 +679,8 @@ def run_decode(
do_print=False,
n_splits_linear=n_splits_linear,
n_splits_down_proj=n_splits_down_proj,
group_size=group_size
group_size=group_size,
asym=asym,
)

dist.barrier()
Expand Down Expand Up @@ -827,6 +844,7 @@ def run_prefill(
layer_indexs = range(layer_start, layer_end)
n_splits_linear = len(model.model.layers[0].mlp.gate_proj_dq_list)
n_splits_down_proj = len(model.model.layers[0].mlp.down_proj_dq_list)
asym = getattr(model.config, "asym", False)
for layer_idx in layer_indexs:
curr_layer = model.model.layers[layer_idx]
attn_layer = curr_layer.self_attn
Expand All @@ -838,10 +856,17 @@ def run_prefill(
mlp_layer.down_proj_dq_list]:
l_weights = []
scales = []
zeros = []
for l in layer_list:
l_weights.append(l.weight)
scales.append(l.scale)
weights.append((torch.stack(l_weights, axis=0), torch.stack(scales, axis=0)))
if l.zero is not None:
zeros.append(l.zero)
if len(zeros):
weights.append((torch.stack(l_weights, axis=0), torch.stack(scales, axis=0),
torch.stack(zeros, axis=0)))
else:
weights.append((torch.stack(l_weights, axis=0), torch.stack(scales, axis=0)))

cached_cos = curr_layer.self_attn.rotary_emb.cos_cached.to(torch.float16)
cached_sin = curr_layer.self_attn.rotary_emb.sin_cached.to(torch.float16)
Expand All @@ -864,7 +889,8 @@ def run_prefill(
transpose_value=transpose_value_cache,
n_splits_linear=n_splits_linear,
n_splits_down_proj=n_splits_down_proj,
group_size=group_size
group_size=group_size,
asym=asym
)

layer_weights.extend(weights)
Expand Down
Loading