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

Divide core-xe packages (#11131)

Browse source 
* temp

* add batch

* fix style

* update package name

* fix style

* add workflow

* use temp version to run uts

* trigger performance test

* trigger win igpu perf

* revert workflow & setup
This commit is contained in:
Yina Chen 2024-05-28 12:00:18 +08:00 committed by GitHub
parent c9168b85b7
commit b6b70d1ba0
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
28 changed files with 427 additions and 373 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

8
.github/actions/llm/setup-llm-env/action.yml vendored
View file

@ -13,12 +13,20 @@ runs:
run: | run: |
# make sure we install the latest version for bigdl-core-xe related packages # make sure we install the latest version for bigdl-core-xe related packages
pip uninstall bigdl-core-xe -y || true pip uninstall bigdl-core-xe -y || true
pip uninstall bigdl-core-xe-batch -y || true
pip uninstall bigdl-core-xe-addons -y || true
pip uninstall bigdl-core-xe-esimd -y || true pip uninstall bigdl-core-xe-esimd -y || true
pip uninstall bigdl-core-xe-21 -y || true pip uninstall bigdl-core-xe-21 -y || true
pip uninstall bigdl-core-xe-batch-21 -y || true
pip uninstall bigdl-core-xe-addons-21 -y || true
pip uninstall bigdl-core-xe-esimd-21 -y || true pip uninstall bigdl-core-xe-esimd-21 -y || true
sed -i 's/"bigdl-core-xe==" + CORE_XE_VERSION + "/"bigdl-core-xe/g' python/llm/setup.py sed -i 's/"bigdl-core-xe==" + CORE_XE_VERSION + "/"bigdl-core-xe/g' python/llm/setup.py
sed -i 's/"bigdl-core-xe-batch==" + CORE_XE_VERSION + "/"bigdl-core-xe-batch/g' python/llm/setup.py
sed -i 's/"bigdl-core-xe-addons==" + CORE_XE_VERSION + "/"bigdl-core-xe-addons/g' python/llm/setup.py
sed -i 's/"bigdl-core-xe-esimd==" + CORE_XE_VERSION + "/"bigdl-core-xe-esimd/g' python/llm/setup.py sed -i 's/"bigdl-core-xe-esimd==" + CORE_XE_VERSION + "/"bigdl-core-xe-esimd/g' python/llm/setup.py
sed -i 's/"bigdl-core-xe-21==" + CORE_XE_VERSION/"bigdl-core-xe-21"/g' python/llm/setup.py sed -i 's/"bigdl-core-xe-21==" + CORE_XE_VERSION/"bigdl-core-xe-21"/g' python/llm/setup.py
sed -i 's/"bigdl-core-xe-batch-21==" + CORE_XE_VERSION/"bigdl-core-xe-batch-21"/g' python/llm/setup.py
sed -i 's/"bigdl-core-xe-addons-21==" + CORE_XE_VERSION/"bigdl-core-xe-addons-21"/g' python/llm/setup.py
sed -i 's/"bigdl-core-xe-esimd-21==" + CORE_XE_VERSION/"bigdl-core-xe-esimd-21"/g' python/llm/setup.py sed -i 's/"bigdl-core-xe-esimd-21==" + CORE_XE_VERSION/"bigdl-core-xe-esimd-21"/g' python/llm/setup.py
pip install requests pip install requests

2
.github/workflows/llm_performance_tests.yml vendored
View file

@ -312,6 +312,8 @@ jobs:
# shell: bash # shell: bash
# run: | # run: |
# sed -i 's/"bigdl-core-xe-21==" + CORE_XE_VERSION/"bigdl-core-xe-21"/g' python/llm/setup.py # sed -i 's/"bigdl-core-xe-21==" + CORE_XE_VERSION/"bigdl-core-xe-21"/g' python/llm/setup.py
# sed -i 's/"bigdl-core-xe-batch-21==" + CORE_XE_VERSION/"bigdl-core-xe-batch-21"/g' python/llm/setup.py
# sed -i 's/"bigdl-core-xe-addons-21==" + CORE_XE_VERSION/"bigdl-core-xe-addons-21"/g' python/llm/setup.py
# sed -i 's/"bigdl-core-xe-esimd-21==" + CORE_XE_VERSION/"bigdl-core-xe-esimd-21"/g' python/llm/setup.py # sed -i 's/"bigdl-core-xe-esimd-21==" + CORE_XE_VERSION/"bigdl-core-xe-esimd-21"/g' python/llm/setup.py
# - name: Install ipex-llm and other related packages (install from source) # - name: Install ipex-llm and other related packages (install from source)

2
python/llm/setup.py
View file

@ -298,6 +298,8 @@ def setup_package():
"torchvision==0.16.0a0", "torchvision==0.16.0a0",
"intel_extension_for_pytorch==2.1.10+xpu", "intel_extension_for_pytorch==2.1.10+xpu",
"bigdl-core-xe-21==" + CORE_XE_VERSION, "bigdl-core-xe-21==" + CORE_XE_VERSION,
"bigdl-core-xe-batch-21==" + CORE_XE_VERSION,
"bigdl-core-xe-addons-21==" + CORE_XE_VERSION,
"bigdl-core-xe-esimd-21==" + CORE_XE_VERSION] "bigdl-core-xe-esimd-21==" + CORE_XE_VERSION]
xpu_21_requires += oneapi_2024_0_requires xpu_21_requires += oneapi_2024_0_requires
# default to ipex 2.1 for linux and windows # default to ipex 2.1 for linux and windows

4
python/llm/src/ipex_llm/transformers/bmm.py
View file

@ -16,7 +16,7 @@
import torch import torch
import linear_q4_0 import xe_linear
torch_bmm_old_ = torch.bmm torch_bmm_old_ = torch.bmm
@ -30,7 +30,7 @@ def torch_bmm(a, b):
if a.size(1) == 1: if a.size(1) == 1:
torch_bmm_old_(a, b, out=C) torch_bmm_old_(a, b, out=C)
else: else:
linear_q4_0.bmm(a.contiguous(), b.contiguous(), C) xe_linear.bmm(a.contiguous(), b.contiguous(), C)
return C return C

4
python/llm/src/ipex_llm/transformers/embedding.py
View file

@ -104,11 +104,11 @@ class LowBitEmbedding(torch.nn.Embedding):
"`LowBitEmbedding` only supports GPU now.") "`LowBitEmbedding` only supports GPU now.")
try: try:
import intel_extension_for_pytorch import intel_extension_for_pytorch
import linear_q4_0 import xe_linear
except ModuleNotFoundError: except ModuleNotFoundError:
invalidInputError(False, invalidInputError(False,
"Please `pip install bigdl_core_xe` first.") "Please `pip install bigdl_core_xe` first.")
result = linear_q4_0.dequantize_rows(x.contiguous(), self.weight.data, result = xe_linear.dequantize_rows(x.contiguous(), self.weight.data,
self.weight.qtype, self.embedding_dim) self.weight.qtype, self.embedding_dim)
return result.to(self.torch_dtype) return result.to(self.torch_dtype)

8
python/llm/src/ipex_llm/transformers/layers/rope_embedding.py
View file

@ -42,9 +42,9 @@ class FastRopeEmbedding(torch.autograd.Function):
@staticmethod @staticmethod
@custom_fwd @custom_fwd
def forward(ctx, x, position_ids): def forward(ctx, x, position_ids):
import linear_q4_0 import xe_addons
x_embed = torch.empty(x.shape, dtype=x.dtype, device=x.device) x_embed = torch.empty(x.shape, dtype=x.dtype, device=x.device)
linear_q4_0.apply_rotary_embedding_half_q_or_k(x, position_ids, xe_addons.apply_rotary_embedding_half_q_or_k(x, position_ids,
x_embed, False) x_embed, False)
ctx.save_for_backward(position_ids) ctx.save_for_backward(position_ids)
return x_embed return x_embed
@ -52,13 +52,13 @@ class FastRopeEmbedding(torch.autograd.Function):
@staticmethod @staticmethod
@custom_bwd @custom_bwd
def backward(ctx, grad_output): def backward(ctx, grad_output):
import linear_q4_0 import xe_addons
# LOG.info(f"backward, grad_output: {grad_output}") # LOG.info(f"backward, grad_output: {grad_output}")
position_ids, = ctx.saved_tensors position_ids, = ctx.saved_tensors
dx = torch.empty(grad_output.shape, dx = torch.empty(grad_output.shape,
dtype=grad_output.dtype, dtype=grad_output.dtype,
device=grad_output.device) device=grad_output.device)
linear_q4_0.apply_rotary_embedding_half_q_or_k(grad_output, xe_addons.apply_rotary_embedding_half_q_or_k(grad_output,
position_ids, position_ids,
dx, dx,
True) True)

50
python/llm/src/ipex_llm/transformers/low_bit_linear.py
View file

@ -320,6 +320,26 @@ def reshape_lm_head_input(x):
return x return x
def use_batch_forward(x: torch.Tensor, qtype: int, output_len: int):
device = get_xpu_device_type(x)
batch_size = x.shape[0]
hard_condition = (
x.dtype in [torch.float, torch.half]
and x.shape[1] % 256 == 0
and output_len % 32 == 0
and device in ["arc", "flex", "pvc", "mtl"]
and qtype in [SYM_INT4, ASYM_INT4, SYM_INT8, FP4,
FP8E5, FP6]
and batch_size <= 64
)
if hard_condition:
return (
batch_size > 1
or (device in ["arc", "flex"] and qtype in [SYM_INT8, FP4])
)
return False
# Rename to FP4Params to trigger initializing # Rename to FP4Params to trigger initializing
# the params layer with all parameters on the CPU # the params layer with all parameters on the CPU
# https://github.com/huggingface/accelerate/blob/main/src/accelerate/utils/modeling.py#L333 # https://github.com/huggingface/accelerate/blob/main/src/accelerate/utils/modeling.py#L333
@ -524,8 +544,8 @@ class MatMulLowBit(torch.autograd.Function):
@custom_fwd @custom_fwd
def forward(ctx, A, weight, input_seq_size): def forward(ctx, A, weight, input_seq_size):
ctx.is_empty = False ctx.is_empty = False
import linear_q4_0 import xe_linear
result = linear_q4_0.forward_new(A, weight.data, weight.qtype, input_seq_size) result = xe_linear.forward_new(A, weight.data, weight.qtype, input_seq_size)
if any(ctx.needs_input_grad[:2]): if any(ctx.needs_input_grad[:2]):
ctx.tensors = (A, weight) ctx.tensors = (A, weight)
else: else:
@ -535,7 +555,7 @@ class MatMulLowBit(torch.autograd.Function):
@staticmethod @staticmethod
@custom_bwd @custom_bwd
def backward(ctx, grad_output): def backward(ctx, grad_output):
import linear_q4_0 import xe_linear
if ctx.is_empty: if ctx.is_empty:
bias_grad = None if ctx.bias is None else torch.zeros_like(ctx.bias) bias_grad = None if ctx.bias is None else torch.zeros_like(ctx.bias)
return torch.zeros_like(ctx.A), torch.zeros_like(ctx.B), None, bias_grad, None return torch.zeros_like(ctx.A), torch.zeros_like(ctx.B), None, bias_grad, None
@ -545,7 +565,7 @@ class MatMulLowBit(torch.autograd.Function):
if req_gradA: if req_gradA:
if torch.xpu.is_autocast_xpu_enabled(): if torch.xpu.is_autocast_xpu_enabled():
grad_output = grad_output.to(torch.xpu.get_autocast_xpu_dtype()) grad_output = grad_output.to(torch.xpu.get_autocast_xpu_dtype())
dequant_weight = linear_q4_0.dequant(A, weight.data, weight.qtype) dequant_weight = xe_linear.dequant(A, weight.data, weight.qtype)
grad_A = torch.matmul(grad_output, dequant_weight.reshape(weight._shape)) grad_A = torch.matmul(grad_output, dequant_weight.reshape(weight._shape))
return grad_A, grad_weight, None return grad_A, grad_weight, None
@ -659,7 +679,7 @@ class LowBitLinear(nn.Linear):
# GPU logic # GPU logic
try: try:
import intel_extension_for_pytorch import intel_extension_for_pytorch
import linear_q4_0 import xe_linear
from ipex_llm.transformers.models.utils import use_xmx from ipex_llm.transformers.models.utils import use_xmx
except ModuleNotFoundError: except ModuleNotFoundError:
invalidInputError(False, invalidInputError(False,
@ -678,12 +698,12 @@ class LowBitLinear(nn.Linear):
if x_2d.requires_grad: if x_2d.requires_grad:
result = MatMulLowBit.apply(x_2d, self.weight, input_seq_size) result = MatMulLowBit.apply(x_2d, self.weight, input_seq_size)
else: else:
result = linear_q4_0.forward_new(x_2d, self.weight.data, result = xe_linear.forward_new(x_2d, self.weight.data,
self.weight.qtype, self.weight.qtype,
input_seq_size) input_seq_size)
elif self.enable_xetla: elif self.enable_xetla:
x_2d = x_2d.half() x_2d = x_2d.half()
result = linear_q4_0.mm_xetla(x_2d, self.weight.data, self.qtype) result = xe_linear.mm_xetla(x_2d, self.weight.data, self.qtype)
else: else:
# inference path # inference path
# current workaround to reduce first token latency of fp32 input # current workaround to reduce first token latency of fp32 input
@ -696,11 +716,23 @@ class LowBitLinear(nn.Linear):
if self.conver_to_half and x_2d.shape[0] > 1 and x_2d.dtype == torch.float32 and \ if self.conver_to_half and x_2d.shape[0] > 1 and x_2d.dtype == torch.float32 and \
not use_xmx(x_2d, self.weight.qtype): not use_xmx(x_2d, self.weight.qtype):
x_2d = x_2d.half() x_2d = x_2d.half()
result = linear_q4_0.forward_new(x_2d, self.weight.data, self.weight.qtype, if use_batch_forward(x_2d, self.weight.qtype, self.out_len):
import xe_batch
result = xe_batch.batch_forward(x_2d, self.weight.data,
self.weight.qtype,
input_seq_size)
else:
result = xe_linear.forward_new(x_2d, self.weight.data, self.weight.qtype,
input_seq_size) input_seq_size)
result = result.to(x.dtype) result = result.to(x.dtype)
else: else:
result = linear_q4_0.forward_new(x_2d, self.weight.data, self.weight.qtype, if use_batch_forward(x_2d, self.weight.qtype, self.out_len):
import xe_batch
result = xe_batch.batch_forward(x_2d, self.weight.data,
self.weight.qtype,
input_seq_size)
else:
result = xe_linear.forward_new(x_2d, self.weight.data, self.weight.qtype,
input_seq_size) input_seq_size)
if do_empty_cache: if do_empty_cache:
torch.xpu.empty_cache() torch.xpu.empty_cache()

18
python/llm/src/ipex_llm/transformers/models/baichuan.py
View file

@ -168,8 +168,8 @@ def baichuan_attention_forward_7b_quantized(
dtype=torch.float32).to(query_states.dtype) dtype=torch.float32).to(query_states.dtype)
attn_output = torch.matmul(attn_weights, value_states) attn_output = torch.matmul(attn_weights, value_states)
else: else:
import linear_q4_0 import xe_addons
attn_output = linear_q4_0.sdp_fp8(query_states, key_states, value_states, attn_output = xe_addons.sdp_fp8(query_states, key_states, value_states,
attention_mask) attention_mask)
attn_weights = None attn_weights = None
@ -277,8 +277,9 @@ def baichuan_attention_forward_7b_origin(
attn_weights = None attn_weights = None
elif not self.training and not hidden_states.requires_grad and \ elif not self.training and not hidden_states.requires_grad and \
use_sdp(q_len, key_states.shape[2], self.head_dim, query_states): use_sdp(q_len, key_states.shape[2], self.head_dim, query_states):
import linear_q4_0 import xe_addons
attn_output = linear_q4_0.sdp(query_states, key_states, value_states, attention_mask) attn_output = xe_addons.sdp(query_states, key_states, value_states,
attention_mask)
attn_output = attn_output.view(query_states.shape) attn_output = attn_output.view(query_states.shape)
attn_weights = None attn_weights = None
else: else:
@ -400,8 +401,8 @@ def baichuan_attention_forward_13b_quantized(
query_states.dtype) query_states.dtype)
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) attn_weights = torch.matmul(query_states, key_states.transpose(2, 3))
else: else:
import linear_q4_0 import xe_addons
attn_weights = linear_q4_0.query_key_fp8_matmul(query_states, key_states) attn_weights = xe_addons.query_key_fp8_matmul(query_states, key_states)
attn_weights = attn_weights / math.sqrt(self.head_dim) attn_weights = attn_weights / math.sqrt(self.head_dim)
@ -419,8 +420,9 @@ def baichuan_attention_forward_13b_quantized(
if query_states.size(2) != 1 or query_states.device.type != 'xpu': if query_states.size(2) != 1 or query_states.device.type != 'xpu':
attn_output = torch.matmul(attn_weights, value_states) attn_output = torch.matmul(attn_weights, value_states)
else: else:
import linear_q4_0 import xe_addons
attn_output = linear_q4_0.attn_value_fp8_matmul(attn_weights, value_states) attn_output = xe_addons.attn_value_fp8_matmul(attn_weights,
value_states)
attn_output = attn_output.transpose(1, 2) attn_output = attn_output.transpose(1, 2)
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size) attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)

34
python/llm/src/ipex_llm/transformers/models/baichuan2.py
View file

@ -41,9 +41,9 @@ def pre_compute_inv_freq(module: torch.nn.Module):
def baichuan_13b_rms_norm_forward(self, hidden_states): def baichuan_13b_rms_norm_forward(self, hidden_states):
if hidden_states.device.type == "xpu" and not (self.training or hidden_states.requires_grad): if hidden_states.device.type == "xpu" and not (self.training or hidden_states.requires_grad):
import linear_q4_0 import xe_addons
x_2d = hidden_states.reshape(-1, hidden_states.size(-1)).contiguous() x_2d = hidden_states.reshape(-1, hidden_states.size(-1)).contiguous()
output = linear_q4_0.rms_norm(self.weight, x_2d, self.epsilon) output = xe_addons.rms_norm(self.weight, x_2d, self.epsilon)
return output.reshape(hidden_states.shape) return output.reshape(hidden_states.shape)
input_dtype = hidden_states.dtype input_dtype = hidden_states.dtype
@ -60,10 +60,10 @@ def baichuan_mlp_forward(
x_2d = x.view(-1, x.shape[-1]) x_2d = x.view(-1, x.shape[-1])
qtype = getattr(self.gate_proj, "qtype", None) qtype = getattr(self.gate_proj, "qtype", None)
if mlp_fusion_check(x_2d, qtype, self.training) and not self.down_proj.enable_xetla: if mlp_fusion_check(x_2d, qtype, self.training) and not self.down_proj.enable_xetla:
import linear_q4_0 import xe_linear
if not x_2d.is_contiguous(): if not x_2d.is_contiguous():
x_2d = x_2d.contiguous() x_2d = x_2d.contiguous()
return self.down_proj(linear_q4_0.mlp_forward_xpu( return self.down_proj(xe_linear.mlp_forward_xpu(
x_2d, self.gate_proj.weight.data, self.up_proj.weight.data, x_2d, self.gate_proj.weight.data, self.up_proj.weight.data,
x_2d.shape[0], x_2d.shape[1], self.gate_proj.out_len, x_2d.shape[0], x_2d.shape[1], self.gate_proj.out_len,
SILU, qtype SILU, qtype
@ -96,8 +96,8 @@ def baichuan_attention_forward_7b(
# IPEX-LLM OPT: fuse rope # IPEX-LLM OPT: fuse rope
if should_use_fuse_rope(hidden_states, position_ids, self.training): if should_use_fuse_rope(hidden_states, position_ids, self.training):
import linear_q4_0 import xe_addons
linear_q4_0.rotary_half_inplaced(self.rotary_emb.inv_freq, position_ids, xe_addons.rotary_half_inplaced(self.rotary_emb.inv_freq, position_ids,
query_states, key_states) query_states, key_states)
else: else:
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len) cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
@ -126,18 +126,20 @@ def baichuan_attention_forward_7b(
value_states.to(dtype=torch.float16), value_states.to(dtype=torch.float16),
is_causal=True).to(hidden_states.dtype) is_causal=True).to(hidden_states.dtype)
elif use_sdp(q_len, kv_seq_len, self.head_dim, query_states): elif use_sdp(q_len, kv_seq_len, self.head_dim, query_states):
import linear_q4_0 import xe_addons
if use_quantize_kv: if use_quantize_kv:
attn_output = linear_q4_0.sdp_fp8(query_states, key_states, value_states, attn_output = xe_addons.sdp_fp8(query_states, key_states, value_states,
attention_mask) attention_mask)
else: else:
attn_output = linear_q4_0.sdp(query_states, key_states, value_states, attention_mask) attn_output = xe_addons.sdp(query_states, key_states, value_states,
attention_mask)
elif use_sdp_causal(q_len, kv_seq_len, self.head_dim, query_states, self.training): elif use_sdp_causal(q_len, kv_seq_len, self.head_dim, query_states, self.training):
import linear_q4_0 import xe_addons
if use_quantize_kv: if use_quantize_kv:
attn_output = linear_q4_0.sdp_fp8_causal(query_states, key_states, value_states) attn_output = xe_addons.sdp_fp8_causal(query_states, key_states,
value_states)
else: else:
attn_output = linear_q4_0.sdp_causal(query_states, key_states, value_states) attn_output = xe_addons.sdp_causal(query_states, key_states, value_states)
else: else:
if use_quantize_kv: if use_quantize_kv:
key_states, value_states = restore_fp8_kv_cache(key_states, value_states, key_states, value_states = restore_fp8_kv_cache(key_states, value_states,
@ -202,8 +204,8 @@ def baichuan_attention_forward_13b(
attention_mask = attention_mask[:, None, -q_len:, :] attention_mask = attention_mask[:, None, -q_len:, :]
if use_quantize_kv and q_len == 1: if use_quantize_kv and q_len == 1:
import linear_q4_0 import xe_addons
attn_weights = linear_q4_0.query_key_fp8_matmul(query_states, key_states) attn_weights = xe_addons.query_key_fp8_matmul(query_states, key_states)
else: else:
if use_quantize_kv: if use_quantize_kv:
key_states, value_states = restore_fp8_kv_cache(key_states, value_states, key_states, value_states = restore_fp8_kv_cache(key_states, value_states,
@ -216,8 +218,8 @@ def baichuan_attention_forward_13b(
attn_weights = attn_weights.to(query_states.dtype) attn_weights = attn_weights.to(query_states.dtype)
attn_weights = torch.nn.functional.softmax(attn_weights, dim=-1) attn_weights = torch.nn.functional.softmax(attn_weights, dim=-1)
if use_quantize_kv and q_len == 1: if use_quantize_kv and q_len == 1:
import linear_q4_0 import xe_addons
attn_output = linear_q4_0.attn_value_fp8_matmul(attn_weights, value_states) attn_output = xe_addons.attn_value_fp8_matmul(attn_weights, value_states)
else: else:
attn_output = torch.matmul(attn_weights.to(dtype=value_states.dtype), value_states) attn_output = torch.matmul(attn_weights.to(dtype=value_states.dtype), value_states)
attn_output = attn_output.transpose(1, 2) attn_output = attn_output.transpose(1, 2)

4
python/llm/src/ipex_llm/transformers/models/bloom.py
View file

@ -66,8 +66,8 @@ def dropout_add(x: torch.Tensor, residual: torch.Tensor, prob: float, training:
def bloom_layer_norm_forward(self, hidden_states): def bloom_layer_norm_forward(self, hidden_states):
if use_fused_layer_norm(hidden_states, self.training): if use_fused_layer_norm(hidden_states, self.training):
import linear_q4_0 import xe_addons
result = linear_q4_0.fused_layer_norm(hidden_states, result = xe_addons.fused_layer_norm(hidden_states,
[self.weight.size(0)], [self.weight.size(0)],
self.weight, self.weight,
self.bias, self.bias,

12
python/llm/src/ipex_llm/transformers/models/chatglm2.py
View file

@ -111,9 +111,9 @@ def should_split_qkv_tensor(query_layer, bsz, n_head, seq_len):
def chatglm_rms_norm_forward(self, hidden_states): def chatglm_rms_norm_forward(self, hidden_states):
if hidden_states.device.type == "xpu" and not (self.training and hidden_states.requires_grad): if hidden_states.device.type == "xpu" and not (self.training and hidden_states.requires_grad):
import linear_q4_0 import xe_addons
x_2d = hidden_states.reshape(-1, hidden_states.size(-1)).contiguous() x_2d = hidden_states.reshape(-1, hidden_states.size(-1)).contiguous()
output = linear_q4_0.rms_norm(self.weight, x_2d, self.eps) output = xe_addons.rms_norm(self.weight, x_2d, self.eps)
return output.reshape(hidden_states.shape) return output.reshape(hidden_states.shape)
input_dtype = hidden_states.dtype input_dtype = hidden_states.dtype
@ -322,8 +322,8 @@ def chatglm2_quantized_attention_forward_8eb45c(
context_layer = torch.matmul(attn.to(value.dtype), value) context_layer = torch.matmul(attn.to(value.dtype), value)
else: else:
key, value = k_cache, v_cache key, value = k_cache, v_cache
import linear_q4_0 import xe_addons
context_layer = linear_q4_0.sdp_fp8(query_layer, key, value, attn_bias) context_layer = xe_addons.sdp_fp8(query_layer, key, value, attn_bias)
# context_layer's shape: [bs, n_head, seq_len, head_dim] -> [seq_len, bs, n_head * head_dim] # context_layer's shape: [bs, n_head, seq_len, head_dim] -> [seq_len, bs, n_head * head_dim]
context_layer = context_layer.permute(2, 0, 1, 3).contiguous().view(seq_len, batch_size, -1) context_layer = context_layer.permute(2, 0, 1, 3).contiguous().view(seq_len, batch_size, -1)
@ -572,8 +572,8 @@ def core_attn_forward_8eb45c(query_layer, key_layer, value_layer, attention_mask
if use_sdp(query_layer.shape[2], key_layer.shape[2], if use_sdp(query_layer.shape[2], key_layer.shape[2],
query_layer.shape[-1], query_layer): query_layer.shape[-1], query_layer):
import linear_q4_0 import xe_addons
attn_output = linear_q4_0.sdp(query_layer, key_layer, value_layer, attn_bias) attn_output = xe_addons.sdp(query_layer, key_layer, value_layer, attn_bias)
context_layer = attn_output.view(query_layer.shape) context_layer = attn_output.view(query_layer.shape)
else: else:
head_dim = query_layer.size(-1) head_dim = query_layer.size(-1)

13
python/llm/src/ipex_llm/transformers/models/cohere.py
View file

@ -261,9 +261,8 @@ def cohere_attention_forward_quantized(
cache_kwargs, new_layout=True) cache_kwargs, new_layout=True)
if q_len == 1 and query_states.device.type == 'xpu' and not self.training \ if q_len == 1 and query_states.device.type == 'xpu' and not self.training \
and not hidden_states.requires_grad: and not hidden_states.requires_grad:
import linear_q4_0 import xe_addons
attn_output = linear_q4_0.sdp_fp8(query_states, key_states, value_states, attn_output = xe_addons.sdp_fp8(query_states, key_states, value_states, attention_mask)
attention_mask)
attn_weights = None attn_weights = None
else: else:
key_states, value_states = restore_fp8_kv_cache(key_states, key_states, value_states = restore_fp8_kv_cache(key_states,
@ -325,8 +324,8 @@ def cohere_attention_forward_origin(
cache_k = past_key_value.key_cache[self.layer_idx] cache_k = past_key_value.key_cache[self.layer_idx]
cache_v = past_key_value.value_cache[self.layer_idx] cache_v = past_key_value.value_cache[self.layer_idx]
kv_seq_len = cache_k.shape[-2] kv_seq_len = cache_k.shape[-2]
import linear_q4_0 import xe_linear
query_states, key_states, value_states = linear_q4_0.forward_qkv(hidden_states, query_states, key_states, value_states = xe_linear.forward_qkv(hidden_states,
self.q_proj.weight, self.q_proj.weight,
self.k_proj.weight, self.k_proj.weight,
self.v_proj.weight, self.v_proj.weight,
@ -421,12 +420,12 @@ def cohere_attention_forward_origin(
attn_weights = None attn_weights = None
elif not self.training and not hidden_states.requires_grad and \ elif not self.training and not hidden_states.requires_grad and \
use_sdp(q_len, key_states.shape[2], self.head_dim, query_states): use_sdp(q_len, key_states.shape[2], self.head_dim, query_states):
import linear_q4_0 import xe_addons
if attention_mask is not None: if attention_mask is not None:
causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
else: else:
causal_mask = None causal_mask = None
attn_output = linear_q4_0.sdp(query_states, key_states, value_states, causal_mask) attn_output = xe_addons.sdp(query_states, key_states, value_states, causal_mask)
attn_output = attn_output.view(query_states.shape) attn_output = attn_output.view(query_states.shape)
attn_weights = None attn_weights = None
else: else:

12
python/llm/src/ipex_llm/transformers/models/gemma.py
View file

@ -79,9 +79,9 @@ def should_use_fuse_rope(self, hidden_states, position_ids):
def gemma_rms_norm_forward(self, hidden_states): def gemma_rms_norm_forward(self, hidden_states):
if hidden_states.device.type == "xpu" and not (self.training and hidden_states.requires_grad): if hidden_states.device.type == "xpu" and not (self.training and hidden_states.requires_grad):
import linear_q4_0 import xe_addons
x_2d = hidden_states.reshape(-1, hidden_states.size(-1)).contiguous() x_2d = hidden_states.reshape(-1, hidden_states.size(-1)).contiguous()
output = linear_q4_0.rms_norm(self.weight + 1, x_2d, self.eps) output = xe_addons.rms_norm(self.weight + 1, x_2d, self.eps)
return output.reshape(hidden_states.shape) return output.reshape(hidden_states.shape)
input_dtype = hidden_states.dtype input_dtype = hidden_states.dtype
@ -100,10 +100,10 @@ def gemma_mlp_forward(
bsz, hidden_size = x_2d.shape bsz, hidden_size = x_2d.shape
qtype = getattr(self.gate_proj, "qtype", None) qtype = getattr(self.gate_proj, "qtype", None)
if mlp_fusion_check(x_2d, qtype, self.training) and not self.down_proj.enable_xetla: if mlp_fusion_check(x_2d, qtype, self.training) and not self.down_proj.enable_xetla:
import linear_q4_0 import xe_linear
if not x_2d.is_contiguous(): if not x_2d.is_contiguous():
x_2d = x_2d.contiguous() x_2d = x_2d.contiguous()
out = self.down_proj(linear_q4_0.mlp_forward_xpu( out = self.down_proj(xe_linear.mlp_forward_xpu(
x_2d, self.gate_proj.weight.data, self.up_proj.weight.data, x_2d, self.gate_proj.weight.data, self.up_proj.weight.data,
x_2d.shape[0], x_2d.shape[1], self.gate_proj.out_len, x_2d.shape[0], x_2d.shape[1], self.gate_proj.out_len,
GELU, qtype GELU, qtype
@ -146,8 +146,8 @@ def gemma_attention_forward(
kv_seq_len = cache_k.shape[-2] kv_seq_len = cache_k.shape[-2]
import linear_q4_0 import xe_linear
query_states, key_states, value_states = linear_q4_0.forward_qkv(hidden_states, query_states, key_states, value_states = xe_linear.forward_qkv(hidden_states,
self.q_proj.weight, self.q_proj.weight,
self.k_proj.weight, self.k_proj.weight,
self.v_proj.weight, self.v_proj.weight,

12
python/llm/src/ipex_llm/transformers/models/internlm.py
View file

@ -398,18 +398,18 @@ def internlm_xcomposser2_attention_forward(
# IPEX-LLM OPT: sdp # IPEX-LLM OPT: sdp
if use_sdp(q_len, kv_seq_len, self.head_dim, query_states): if use_sdp(q_len, kv_seq_len, self.head_dim, query_states):
import linear_q4_0 import xe_linear
if use_quantize_kv: if use_quantize_kv:
attn_output = linear_q4_0.sdp_fp8(query_states, key_states, value_states, attn_output = xe_linear.sdp_fp8(query_states, key_states, value_states,
attention_mask) attention_mask)
else: else:
attn_output = linear_q4_0.sdp(query_states, key_states, value_states, attention_mask) attn_output = xe_linear.sdp(query_states, key_states, value_states, attention_mask)
elif use_sdp_causal(q_len, kv_seq_len, self.head_dim, query_states, self.training): elif use_sdp_causal(q_len, kv_seq_len, self.head_dim, query_states, self.training):
import linear_q4_0 import xe_linear
if use_quantize_kv: if use_quantize_kv:
attn_output = linear_q4_0.sdp_fp8_causal(query_states, key_states, value_states) attn_output = xe_linear.sdp_fp8_causal(query_states, key_states, value_states)
else: else:
attn_output = linear_q4_0.sdp_causal(query_states, key_states, value_states) attn_output = xe_linear.sdp_causal(query_states, key_states, value_states)
else: else:
if use_quantize_kv: if use_quantize_kv:
key_states, value_states = restore_fp8_kv_cache(key_states, value_states, key_states, value_states = restore_fp8_kv_cache(key_states, value_states,

55
python/llm/src/ipex_llm/transformers/models/llama.py
View file

@ -169,9 +169,9 @@ def llama_model_forward_4_38(
def llama_rms_norm_forward(self, hidden_states): def llama_rms_norm_forward(self, hidden_states):
if hidden_states.device.type == "xpu" and not (self.training and hidden_states.requires_grad): if hidden_states.device.type == "xpu" and not (self.training and hidden_states.requires_grad):
import linear_q4_0 import xe_addons
x_2d = hidden_states.reshape(-1, hidden_states.size(-1)).contiguous() x_2d = hidden_states.reshape(-1, hidden_states.size(-1)).contiguous()
output = linear_q4_0.rms_norm(self.weight, x_2d, self.variance_epsilon) output = xe_addons.rms_norm(self.weight, x_2d, self.variance_epsilon)
return output.reshape(hidden_states.shape) return output.reshape(hidden_states.shape)
input_dtype = hidden_states.dtype input_dtype = hidden_states.dtype
@ -190,10 +190,10 @@ def llama_mlp_forward(
bsz, hidden_size = x_2d.shape bsz, hidden_size = x_2d.shape
qtype = getattr(self.gate_proj, "qtype", None) qtype = getattr(self.gate_proj, "qtype", None)
if mlp_fusion_check(x_2d, qtype, self.training) and not self.down_proj.enable_xetla: if mlp_fusion_check(x_2d, qtype, self.training) and not self.down_proj.enable_xetla:
import linear_q4_0 import xe_linear
if not x_2d.is_contiguous(): if not x_2d.is_contiguous():
x_2d = x_2d.contiguous() x_2d = x_2d.contiguous()
out = self.down_proj(linear_q4_0.mlp_forward_xpu( out = self.down_proj(xe_linear.mlp_forward_xpu(
x_2d, self.gate_proj.weight.data, self.up_proj.weight.data, x_2d, self.gate_proj.weight.data, self.up_proj.weight.data,
x_2d.shape[0], x_2d.shape[1], self.gate_proj.out_len, x_2d.shape[0], x_2d.shape[1], self.gate_proj.out_len,
SILU, qtype SILU, qtype
@ -429,8 +429,8 @@ def llama_attention_forward_4_31_quantized(
dtype=hidden_states.dtype, dtype=hidden_states.dtype,
device=device device=device
) )
import linear_q4_0 import xe_linear
query_states, key_states, value_states = linear_q4_0.forward_qkv(hidden_states, query_states, key_states, value_states = xe_linear.forward_qkv(hidden_states,
self.q_proj.weight, self.q_proj.weight,
self.k_proj.weight, self.k_proj.weight,
self.v_proj.weight, self.v_proj.weight,
@ -504,9 +504,8 @@ def llama_attention_forward_4_31_quantized(
bsz, q_len, kv_seq_len, bsz, q_len, kv_seq_len,
self.head_dim, self.num_heads, output_attentions) self.head_dim, self.num_heads, output_attentions)
else: else:
import linear_q4_0 import xe_addons
attn_output = linear_q4_0.sdp_fp8(query_states, key_states, value_states, attn_output = xe_addons.sdp_fp8(query_states, key_states, value_states, attention_mask)
attention_mask)
attn_weights = None attn_weights = None
attn_output = attn_output.transpose(1, 2).contiguous() attn_output = attn_output.transpose(1, 2).contiguous()
@ -562,8 +561,8 @@ def llama_attention_forward_4_31_original(
kv_seq_len = past_key_value[0].shape[-2] kv_seq_len = past_key_value[0].shape[-2]
cache_k = past_key_value[0] cache_k = past_key_value[0]
cache_v = past_key_value[1] cache_v = past_key_value[1]
import linear_q4_0 import xe_linear
query_states, key_states, value_states = linear_q4_0.forward_qkv(hidden_states, query_states, key_states, value_states = xe_linear.forward_qkv(hidden_states,
self.q_proj.weight, self.q_proj.weight,
self.k_proj.weight, self.k_proj.weight,
self.v_proj.weight, self.v_proj.weight,
@ -625,11 +624,11 @@ def llama_attention_forward_4_31_original(
self.k_proj, self.k_proj,
self.v_proj, self.v_proj,
self.q_proj.weight.qtype,) self.q_proj.weight.qtype,)
import linear_q4_0 import xe_linear
q_out_len = self.q_proj.out_len q_out_len = self.q_proj.out_len
k_out_len = self.k_proj.out_len k_out_len = self.k_proj.out_len
v_out_len = self.v_proj.out_len v_out_len = self.v_proj.out_len
qkv_states = linear_q4_0.mm_xetla(hidden_states, self.qkv_proj_qweight, qkv_states = xe_linear.mm_xetla(hidden_states, self.qkv_proj_qweight,
self.q_proj.weight.qtype) self.q_proj.weight.qtype)
query_states = qkv_states[:, :, :q_out_len] query_states = qkv_states[:, :, :q_out_len]
key_states = qkv_states[:, :, q_out_len:q_out_len + k_out_len] key_states = qkv_states[:, :, q_out_len:q_out_len + k_out_len]
@ -712,8 +711,8 @@ def llama_attention_forward_4_31_original(
attn_weights = None attn_weights = None
elif not self.training and not hidden_states.requires_grad and \ elif not self.training and not hidden_states.requires_grad and \
use_sdp(q_len, key_states.shape[2], self.head_dim, query_states): use_sdp(q_len, key_states.shape[2], self.head_dim, query_states):
import linear_q4_0 import xe_addons
attn_output = linear_q4_0.sdp(query_states, key_states, value_states, attention_mask) attn_output = xe_addons.sdp(query_states, key_states, value_states, attention_mask)
attn_output = attn_output.view(query_states.shape) attn_output = attn_output.view(query_states.shape)
attn_weights = None attn_weights = None
else: else:
@ -811,8 +810,8 @@ def llama_attention_selective_batching_forward_4_31(
past_k = new_cache_k past_k = new_cache_k
past_v = new_cache_v past_v = new_cache_v
hidden_states = hidden_states.view(1, -1) hidden_states = hidden_states.view(1, -1)
import linear_q4_0 import xe_linear
query_states, key_states, value_states = linear_q4_0.forward_qkv(hidden_states, query_states, key_states, value_states = xe_linear.forward_qkv(hidden_states,
self.q_proj.weight, self.q_proj.weight,
self.k_proj.weight, self.k_proj.weight,
self.v_proj.weight, self.v_proj.weight,
@ -1028,8 +1027,8 @@ def llama_attention_forward_4_38_quantized(
dtype=hidden_states.dtype, dtype=hidden_states.dtype,
device=device device=device
) )
import linear_q4_0 import xe_linear
query_states, key_states, value_states = linear_q4_0.forward_qkv(hidden_states, query_states, key_states, value_states = xe_linear.forward_qkv(hidden_states,
self.q_proj.weight, self.q_proj.weight,
self.k_proj.weight, self.k_proj.weight,
self.v_proj.weight, self.v_proj.weight,
@ -1176,13 +1175,12 @@ def llama_attention_forward_4_38_quantized(
dtype=torch.float32).to(query_states.dtype) dtype=torch.float32).to(query_states.dtype)
attn_output = torch.matmul(attn_weights, value_states) attn_output = torch.matmul(attn_weights, value_states)
else: else:
import linear_q4_0 import xe_addons
if cache_position is not None: if cache_position is not None:
new_attn_mask = attention_mask[:, :, kv_seq_len-q_len:kv_seq_len, 0:kv_seq_len] new_attn_mask = attention_mask[:, :, kv_seq_len-q_len:kv_seq_len, 0:kv_seq_len]
else: else:
new_attn_mask = attention_mask new_attn_mask = attention_mask
attn_output = linear_q4_0.sdp_fp8(query_states, key_states, value_states, attn_output = xe_addons.sdp_fp8(query_states, key_states, value_states, new_attn_mask)
new_attn_mask)
attn_weights = None attn_weights = None
if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim): if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
@ -1251,8 +1249,8 @@ def llama_attention_forward_4_38_original(
cache_k = past_key_value.key_cache[self.layer_idx] cache_k = past_key_value.key_cache[self.layer_idx]
cache_v = past_key_value.value_cache[self.layer_idx] cache_v = past_key_value.value_cache[self.layer_idx]
kv_seq_len = cache_k.shape[-2] kv_seq_len = cache_k.shape[-2]
import linear_q4_0 import xe_linear
query_states, key_states, value_states = linear_q4_0.forward_qkv(hidden_states, query_states, key_states, value_states = xe_linear.forward_qkv(hidden_states,
self.q_proj.weight, self.q_proj.weight,
self.k_proj.weight, self.k_proj.weight,
self.v_proj.weight, self.v_proj.weight,
@ -1319,11 +1317,11 @@ def llama_attention_forward_4_38_original(
self.k_proj, self.k_proj,
self.v_proj, self.v_proj,
self.q_proj.weight.qtype,) self.q_proj.weight.qtype,)
import linear_q4_0 import xe_linear
q_out_len = self.q_proj.out_len q_out_len = self.q_proj.out_len
k_out_len = self.k_proj.out_len k_out_len = self.k_proj.out_len
v_out_len = self.v_proj.out_len v_out_len = self.v_proj.out_len
qkv_states = linear_q4_0.mm_xetla(hidden_states, qkv_states = xe_linear.mm_xetla(hidden_states,
self.qkv_proj_qweight, self.qkv_proj_qweight,
self.q_proj.weight.qtype) self.q_proj.weight.qtype)
query_states = qkv_states[:, :, :q_out_len] query_states = qkv_states[:, :, :q_out_len]
@ -1425,8 +1423,9 @@ def llama_attention_forward_4_38_original(
attn_weights = None attn_weights = None
elif not self.training and not hidden_states.requires_grad and \ elif not self.training and not hidden_states.requires_grad and \
use_sdp(q_len, key_states.shape[2], self.head_dim, query_states): use_sdp(q_len, key_states.shape[2], self.head_dim, query_states):
import linear_q4_0 import xe_addons
attn_output = linear_q4_0.sdp(query_states, key_states, value_states, new_attention_mask) attn_output = xe_addons.sdp(query_states, key_states, value_states,
new_attention_mask)
attn_output = attn_output.view(query_states.shape) attn_output = attn_output.view(query_states.shape)
attn_weights = None attn_weights = None
else: else:

41
python/llm/src/ipex_llm/transformers/models/mistral.py
View file

@ -278,8 +278,8 @@ def mistral_attention_forward_quantized(
dtype=hidden_states.dtype, dtype=hidden_states.dtype,
device=device device=device
) )
import linear_q4_0 import xe_linear
query_states, key_states, value_states = linear_q4_0.forward_qkv(hidden_states, query_states, key_states, value_states = xe_linear.forward_qkv(hidden_states,
self.q_proj.weight, self.q_proj.weight,
self.k_proj.weight, self.k_proj.weight,
self.v_proj.weight, self.v_proj.weight,
@ -427,8 +427,8 @@ def mistral_attention_forward_quantized(
attn_output = torch.matmul(attn_weights, value_states) attn_output = torch.matmul(attn_weights, value_states)
else: else:
import linear_q4_0 import xe_addons
attn_output = linear_q4_0.sdp_fp8(query_states, key_states, value_states, attn_output = xe_addons.sdp_fp8(query_states, key_states, value_states,
attention_mask) attention_mask)
attn_weights = None attn_weights = None
@ -476,8 +476,8 @@ def mistral_attention_forward_original(
kv_seq_len = past_key_value[0].shape[-2] kv_seq_len = past_key_value[0].shape[-2]
cache_k = past_key_value[0] cache_k = past_key_value[0]
cache_v = past_key_value[1] cache_v = past_key_value[1]
import linear_q4_0 import xe_linear
query_states, key_states, value_states = linear_q4_0.forward_qkv(hidden_states, query_states, key_states, value_states = xe_linear.forward_qkv(hidden_states,
self.q_proj.weight, self.q_proj.weight,
self.k_proj.weight, self.k_proj.weight,
self.v_proj.weight, self.v_proj.weight,
@ -496,11 +496,11 @@ def mistral_attention_forward_original(
self.k_proj, self.k_proj,
self.v_proj, self.v_proj,
self.q_proj.qtype) self.q_proj.qtype)
import linear_q4_0 import xe_linear
q_out_len = self.q_proj.out_len q_out_len = self.q_proj.out_len
k_out_len = self.k_proj.out_len k_out_len = self.k_proj.out_len
v_out_len = self.v_proj.out_len v_out_len = self.v_proj.out_len
qkv_states = linear_q4_0.mm_xetla(hidden_states, qkv_states = xe_linear.mm_xetla(hidden_states,
self.qkv_proj_qweight, self.qkv_proj_qweight,
self.q_proj.qtype) self.q_proj.qtype)
query_states = qkv_states[:, :, :q_out_len] query_states = qkv_states[:, :, :q_out_len]
@ -592,8 +592,8 @@ def mistral_attention_forward_original(
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size) attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
elif use_sdp(q_len, key_states.shape[2], self.head_dim, query_states): elif use_sdp(q_len, key_states.shape[2], self.head_dim, query_states):
# new fp16 sdp doesn't require repeat_kv # new fp16 sdp doesn't require repeat_kv
import linear_q4_0 import xe_addons
attn_output = linear_q4_0.sdp(query_states, key_states, value_states, attention_mask) attn_output = xe_addons.sdp(query_states, key_states, value_states, attention_mask)
attn_output = attn_output.view(query_states.shape) attn_output = attn_output.view(query_states.shape)
attn_weights = None attn_weights = None
attn_output = attn_output.transpose(1, 2).contiguous() attn_output = attn_output.transpose(1, 2).contiguous()
@ -695,8 +695,8 @@ def mistral_attention_forward_4_36_quantized(
dtype=hidden_states.dtype, dtype=hidden_states.dtype,
device=device device=device
) )
import linear_q4_0 import xe_linear
query_states, key_states, value_states = linear_q4_0.forward_qkv(hidden_states, query_states, key_states, value_states = xe_linear.forward_qkv(hidden_states,
self.q_proj.weight, self.q_proj.weight,
self.k_proj.weight, self.k_proj.weight,
self.v_proj.weight, self.v_proj.weight,
@ -852,9 +852,8 @@ def mistral_attention_forward_4_36_quantized(
attn_output = torch.matmul(attn_weights, value_states) attn_output = torch.matmul(attn_weights, value_states)
else: else:
import linear_q4_0 import xe_addons
attn_output = linear_q4_0.sdp_fp8(query_states, key_states, value_states, attn_output = xe_addons.sdp_fp8(query_states, key_states, value_states, attention_mask)
attention_mask)
attn_weights = None attn_weights = None
attn_output_size = (bsz, self.num_heads, q_len, self.head_dim) attn_output_size = (bsz, self.num_heads, q_len, self.head_dim)
@ -905,8 +904,8 @@ def mistral_attention_forward_4_36_original(
kv_seq_len = cache_k.shape[-2] kv_seq_len = cache_k.shape[-2]
import linear_q4_0 import xe_linear
query_states, key_states, value_states = linear_q4_0.forward_qkv(hidden_states, query_states, key_states, value_states = xe_linear.forward_qkv(hidden_states,
self.q_proj.weight, self.q_proj.weight,
self.k_proj.weight, self.k_proj.weight,
self.v_proj.weight, self.v_proj.weight,
@ -931,11 +930,11 @@ def mistral_attention_forward_4_36_original(
self.k_proj, self.k_proj,
self.v_proj, self.v_proj,
self.q_proj.qtype) self.q_proj.qtype)
import linear_q4_0 import xe_linear
q_out_len = self.q_proj.out_len q_out_len = self.q_proj.out_len
k_out_len = self.k_proj.out_len k_out_len = self.k_proj.out_len
v_out_len = self.v_proj.out_len v_out_len = self.v_proj.out_len
qkv_states = linear_q4_0.mm_xetla(hidden_states, qkv_states = xe_linear.mm_xetla(hidden_states,
self.qkv_proj_qweight, self.qkv_proj_qweight,
self.q_proj.qtype) self.q_proj.qtype)
query_states = qkv_states[:, :, :q_out_len] query_states = qkv_states[:, :, :q_out_len]
@ -1033,8 +1032,8 @@ def mistral_attention_forward_4_36_original(
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size) attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
elif use_sdp(q_len, key_states.shape[2], self.head_dim, query_states): elif use_sdp(q_len, key_states.shape[2], self.head_dim, query_states):
# new fp16 sdp doesn't require repeat_kv # new fp16 sdp doesn't require repeat_kv
import linear_q4_0 import xe_addons
attn_output = linear_q4_0.sdp(query_states, key_states, value_states, attention_mask) attn_output = xe_addons.sdp(query_states, key_states, value_states, attention_mask)
attn_output = attn_output.view(query_states.shape) attn_output = attn_output.view(query_states.shape)
attn_weights = None attn_weights = None
attn_output = attn_output.transpose(1, 2).contiguous() attn_output = attn_output.transpose(1, 2).contiguous()

20
python/llm/src/ipex_llm/transformers/models/mixtral.py
View file

@ -105,8 +105,8 @@ def mixtral_moeblock_forward(self,
elif bs < 256 and hidden_states.device.type == 'xpu': elif bs < 256 and hidden_states.device.type == 'xpu':
final_hidden_states = torch.zeros((batch_size * sequence_length, hidden_dim), final_hidden_states = torch.zeros((batch_size * sequence_length, hidden_dim),
dtype=hidden_states.dtype, device=hidden_states.device) dtype=hidden_states.dtype, device=hidden_states.device)
import linear_q4_0 import xe_linear
indexes = linear_q4_0.get_moe_indexes(selected_experts.to(torch.int32).cpu(), 8) indexes = xe_linear.get_moe_indexes(selected_experts.to(torch.int32).cpu(), 8)
for expert_idx in range(self.num_experts): for expert_idx in range(self.num_experts):
expert_layer = self.experts[expert_idx] expert_layer = self.experts[expert_idx]
idx_list = indexes[0][expert_idx] idx_list = indexes[0][expert_idx]
@ -184,8 +184,8 @@ def mixtral_attention_forward(
cache_k = past_key_value.key_cache[self.layer_idx] cache_k = past_key_value.key_cache[self.layer_idx]
cache_v = past_key_value.value_cache[self.layer_idx] cache_v = past_key_value.value_cache[self.layer_idx]
kv_seq_len = cache_k.shape[-2] kv_seq_len = cache_k.shape[-2]
import linear_q4_0 import xe_linear
query_states, key_states, value_states = linear_q4_0.forward_qkv(hidden_states, query_states, key_states, value_states = xe_linear.forward_qkv(hidden_states,
self.q_proj.weight, self.q_proj.weight,
self.k_proj.weight, self.k_proj.weight,
self.v_proj.weight, self.v_proj.weight,
@ -209,8 +209,8 @@ def mixtral_attention_forward(
# cache_k = past_key_value.key_cache[self.layer_idx] # cache_k = past_key_value.key_cache[self.layer_idx]
# cache_v = past_key_value.value_cache[self.layer_idx] # cache_v = past_key_value.value_cache[self.layer_idx]
# kv_seq_len = cache_k.shape[-2] # kv_seq_len = cache_k.shape[-2]
# import linear_q4_0 # import xe_linear
# query_states, key_states = linear_q4_0.forward_qk(hidden_states, # query_states, key_states = xe_linear.forward_qk(hidden_states,
# self.q_proj.weight, # self.q_proj.weight,
# self.k_proj.weight, # self.k_proj.weight,
# position_ids, # position_ids,
@ -333,8 +333,8 @@ def mixtral_attention_forward(
is_causal=True) is_causal=True)
attn_weights = None attn_weights = None
elif use_sdp(query_states.shape[2], key_states.shape[2], self.head_dim, query_states): elif use_sdp(query_states.shape[2], key_states.shape[2], self.head_dim, query_states):
import linear_q4_0 import xe_addons
attn_output = linear_q4_0.sdp(query_states, key_states, value_states, attention_mask) attn_output = xe_addons.sdp(query_states, key_states, value_states, attention_mask)
attn_output = attn_output.view(query_states.shape) attn_output = attn_output.view(query_states.shape)
attn_weights = None attn_weights = None
else: else:
@ -389,8 +389,8 @@ def mixtral_mlp_forward(
) -> torch.Tensor: ) -> torch.Tensor:
qtype = getattr(self.w1, "qtype", None) qtype = getattr(self.w1, "qtype", None)
if mlp_fusion_check(x, qtype, self.training) and not self.w1.enable_xetla: if mlp_fusion_check(x, qtype, self.training) and not self.w1.enable_xetla:
import linear_q4_0 import xe_linear
return self.w2(linear_q4_0.mlp_forward_xpu( return self.w2(xe_linear.mlp_forward_xpu(
x, self.w1.weight.data, self.w3.weight.data, x, self.w1.weight.data, self.w3.weight.data,
x.shape[0], x.shape[1], self.w1.out_len, x.shape[0], x.shape[1], self.w1.out_len,
SILU, qtype, SILU, qtype,

33
python/llm/src/ipex_llm/transformers/models/phi3.py
View file

@ -108,17 +108,17 @@ def attention_forward(
# IPEX-LLM OPT: fuse rope # IPEX-LLM OPT: fuse rope
if should_use_fuse_rope(hidden_states, position_ids, self.training): if should_use_fuse_rope(hidden_states, position_ids, self.training):
import linear_q4_0 import xe_addons
if self.rotary_emb.__class__.__name__ == "Phi3RotaryEmbedding": # 4k if self.rotary_emb.__class__.__name__ == "Phi3RotaryEmbedding": # 4k
linear_q4_0.rotary_half_inplaced(self.rotary_emb.inv_freq, position_ids, xe_addons.rotary_half_inplaced(self.rotary_emb.inv_freq, position_ids,
query_states, key_states) query_states, key_states)
else: # 128k else: # 128k
if kv_seq_len > self.rotary_emb.original_max_position_embeddings: if kv_seq_len > self.rotary_emb.original_max_position_embeddings:
linear_q4_0.rotary_half_inplaced(self.rotary_emb.long_inv_freq, position_ids, xe_addons.rotary_half_inplaced(self.rotary_emb.long_inv_freq,
query_states, key_states) position_ids, query_states, key_states)
else: else:
linear_q4_0.rotary_half_inplaced(self.rotary_emb.inv_freq, position_ids, xe_addons.rotary_half_inplaced(self.rotary_emb.inv_freq,
query_states, key_states) position_ids, query_states, key_states)
# todo: fuse scaling_factor # todo: fuse scaling_factor
query_states *= self.rotary_emb.scaling_factor query_states *= self.rotary_emb.scaling_factor
key_states *= self.rotary_emb.scaling_factor key_states *= self.rotary_emb.scaling_factor
@ -132,18 +132,19 @@ def attention_forward(
self.layer_idx, None) self.layer_idx, None)
if use_sdp(q_len, kv_seq_len, self.head_dim, query_states): if use_sdp(q_len, kv_seq_len, self.head_dim, query_states):
import linear_q4_0 import xe_addons
if isinstance(past_key_value, DynamicFp8Cache): if isinstance(past_key_value, DynamicFp8Cache):
attn_output = linear_q4_0.sdp_fp8(query_states, key_states, value_states, attn_output = xe_addons.sdp_fp8(query_states, key_states, value_states,
attention_mask) attention_mask)
else: else:
attn_output = linear_q4_0.sdp(query_states, key_states, value_states, attention_mask) attn_output = xe_addons.sdp(query_states, key_states, value_states,
attention_mask)
elif use_sdp_causal(q_len, kv_seq_len, self.head_dim, query_states, self.training): elif use_sdp_causal(q_len, kv_seq_len, self.head_dim, query_states, self.training):
import linear_q4_0 import xe_addons
if isinstance(past_key_value, DynamicFp8Cache): if isinstance(past_key_value, DynamicFp8Cache):
attn_output = linear_q4_0.sdp_fp8_causal(query_states, key_states, value_states) attn_output = xe_addons.sdp_fp8_causal(query_states, key_states, value_states)
else: else:
attn_output = linear_q4_0.sdp_causal(query_states, key_states, value_states) attn_output = xe_addons.sdp_causal(query_states, key_states, value_states)
else: else:
if isinstance(past_key_value, DynamicFp8Cache): if isinstance(past_key_value, DynamicFp8Cache):
key_states, value_states = restore_fp8_kv_cache(key_states, value_states, key_states, value_states = restore_fp8_kv_cache(key_states, value_states,
@ -204,8 +205,8 @@ def mlp_forward(
qtype = getattr(self.gate_proj, "qtype", None) qtype = getattr(self.gate_proj, "qtype", None)
if mlp_fusion_check(x_2d, qtype, self.training): if mlp_fusion_check(x_2d, qtype, self.training):
x_2d = x_2d.contiguous() x_2d = x_2d.contiguous()
import linear_q4_0 import xe_linear
return self.down_proj(linear_q4_0.mlp_forward_xpu( return self.down_proj(xe_linear.mlp_forward_xpu(
x_2d, self.gate_proj.weight.data, self.up_proj.weight.data, x_2d, self.gate_proj.weight.data, self.up_proj.weight.data,
x_2d.shape[0], x_2d.shape[1], self.gate_proj.out_features, x_2d.shape[0], x_2d.shape[1], self.gate_proj.out_features,
SILU, qtype SILU, qtype
@ -293,9 +294,9 @@ def phi3v_model_forward_wrapper(origin_model_forward):
def phi3_rms_norm_forward(self, hidden_states): def phi3_rms_norm_forward(self, hidden_states):
if hidden_states.device.type == "xpu" and not (self.training and hidden_states.requires_grad): if hidden_states.device.type == "xpu" and not (self.training and hidden_states.requires_grad):
import linear_q4_0 import xe_addons
x_2d = hidden_states.reshape(-1, hidden_states.size(-1)).contiguous() x_2d = hidden_states.reshape(-1, hidden_states.size(-1)).contiguous()
output = linear_q4_0.rms_norm(self.weight, x_2d, self.variance_epsilon) output = xe_addons.rms_norm(self.weight, x_2d, self.variance_epsilon)
return output.reshape(hidden_states.shape) return output.reshape(hidden_states.shape)
input_dtype = hidden_states.dtype input_dtype = hidden_states.dtype

39
python/llm/src/ipex_llm/transformers/models/qwen.py
View file

@ -92,9 +92,10 @@ def qwen_attention_forward(
rotary_pos_emb = rotary_pos_emb_list[0] rotary_pos_emb = rotary_pos_emb_list[0]
if use_fuse_rope: if use_fuse_rope:
rot_dim = rotary_pos_emb[0].size(-1) rot_dim = rotary_pos_emb[0].size(-1)
import linear_q4_0 import xe_addons
linear_q4_0.rotary_half_inplaced(inv_freq, position_ids, xe_addons.rotary_half_inplaced(inv_freq, position_ids,
query_states[..., :rot_dim], key_states[..., :rot_dim]) query_states[..., :rot_dim],
key_states[..., :rot_dim])
else: else:
rotary_pos_emb = [i[:, -q_len:, :, :].transpose(1, 2) for i in rotary_pos_emb] rotary_pos_emb = [i[:, -q_len:, :, :].transpose(1, 2) for i in rotary_pos_emb]
query_states = apply_rotary_pos_emb(query_states, rotary_pos_emb) query_states = apply_rotary_pos_emb(query_states, rotary_pos_emb)
@ -124,11 +125,11 @@ def qwen_attention_forward(
value_states.to(dtype=torch.float16), value_states.to(dtype=torch.float16),
is_causal=True).to(hidden_states.dtype) is_causal=True).to(hidden_states.dtype)
elif use_sdp_causal(q_len, kv_seq_len, self.head_dim, query_states, self.training): elif use_sdp_causal(q_len, kv_seq_len, self.head_dim, query_states, self.training):
import linear_q4_0 import xe_addons
if use_quantize_kv: if use_quantize_kv:
attn_output = linear_q4_0.sdp_fp8_causal(query_states, key_states, value_states) attn_output = xe_addons.sdp_fp8_causal(query_states, key_states, value_states)
else: else:
attn_output = linear_q4_0.sdp_causal(query_states, key_states, value_states) attn_output = xe_addons.sdp_causal(query_states, key_states, value_states)
else: else:
if q_len > 1: if q_len > 1:
causal_mask = torch.tril( causal_mask = torch.tril(
@ -146,12 +147,12 @@ def qwen_attention_forward(
attention_mask = None attention_mask = None
if use_sdp(q_len, kv_seq_len, self.head_dim, query_states): if use_sdp(q_len, kv_seq_len, self.head_dim, query_states):
import linear_q4_0 import xe_addons
if use_quantize_kv: if use_quantize_kv:
attn_output = linear_q4_0.sdp_fp8(query_states, key_states, value_states, attn_output = xe_addons.sdp_fp8(query_states, key_states, value_states,
attention_mask) attention_mask)
else: else:
attn_output = linear_q4_0.sdp(query_states, key_states, value_states, attn_output = xe_addons.sdp(query_states, key_states, value_states,
attention_mask) attention_mask)
else: else:
if use_quantize_kv: if use_quantize_kv:
@ -221,8 +222,8 @@ def qwen_attention_forward_registered(
rotary_pos_emb = rotary_pos_emb_list[0] rotary_pos_emb = rotary_pos_emb_list[0]
if use_fuse_rope: if use_fuse_rope:
rot_dim = rotary_pos_emb[0].size(-1) rot_dim = rotary_pos_emb[0].size(-1)
import linear_q4_0 import xe_linear
linear_q4_0.rotary_half_inplaced(inv_freq, position_ids, xe_linear.rotary_half_inplaced(inv_freq, position_ids,
query_states[..., :rot_dim], key_states[..., :rot_dim]) query_states[..., :rot_dim], key_states[..., :rot_dim])
else: else:
rotary_pos_emb = [i[:, -q_len:, :, :].transpose(1, 2) for i in rotary_pos_emb] rotary_pos_emb = [i[:, -q_len:, :, :].transpose(1, 2) for i in rotary_pos_emb]
@ -253,11 +254,11 @@ def qwen_attention_forward_registered(
value_states.to(dtype=torch.float16), value_states.to(dtype=torch.float16),
is_causal=True).to(hidden_states.dtype) is_causal=True).to(hidden_states.dtype)
elif use_sdp_causal(q_len, kv_seq_len, self.head_dim, query_states, self.training): elif use_sdp_causal(q_len, kv_seq_len, self.head_dim, query_states, self.training):
import linear_q4_0 import xe_linear
if use_quantize_kv: if use_quantize_kv:
attn_output = linear_q4_0.sdp_fp8_causal(query_states, key_states, value_states) attn_output = xe_linear.sdp_fp8_causal(query_states, key_states, value_states)
else: else:
attn_output = linear_q4_0.sdp_causal(query_states, key_states, value_states) attn_output = xe_linear.sdp_causal(query_states, key_states, value_states)
else: else:
if q_len > 1: if q_len > 1:
causal_mask = registered_causal_mask[ causal_mask = registered_causal_mask[
@ -272,12 +273,12 @@ def qwen_attention_forward_registered(
attention_mask = None attention_mask = None
if use_sdp(q_len, kv_seq_len, self.head_dim, query_states): if use_sdp(q_len, kv_seq_len, self.head_dim, query_states):
import linear_q4_0 import xe_linear
if use_quantize_kv: if use_quantize_kv:
attn_output = linear_q4_0.sdp_fp8(query_states, key_states, value_states, attn_output = xe_linear.sdp_fp8(query_states, key_states, value_states,
attention_mask) attention_mask)
else: else:
attn_output = linear_q4_0.sdp(query_states, key_states, value_states, attn_output = xe_linear.sdp(query_states, key_states, value_states,
attention_mask) attention_mask)
else: else:
if use_quantize_kv: if use_quantize_kv:
@ -310,10 +311,10 @@ def qwen_mlp_forward(self, x: torch.Tensor) -> torch.Tensor:
x_2d = x.view(-1, x.shape[-1]) x_2d = x.view(-1, x.shape[-1])
qtype = getattr(self.w1, "qtype", None) qtype = getattr(self.w1, "qtype", None)
if mlp_fusion_check(x_2d, qtype, self.training) and not self.w1.enable_xetla: if mlp_fusion_check(x_2d, qtype, self.training) and not self.w1.enable_xetla:
import linear_q4_0 import xe_linear
if not x_2d.is_contiguous(): if not x_2d.is_contiguous():
x_2d = x_2d.contiguous() x_2d = x_2d.contiguous()
return self.c_proj(linear_q4_0.mlp_forward_xpu( return self.c_proj(xe_linear.mlp_forward_xpu(
x_2d, self.w2.weight.data, self.w1.weight.data, x_2d, self.w2.weight.data, self.w1.weight.data,
x_2d.shape[0], x_2d.shape[1], self.w2.out_len, x_2d.shape[0], x_2d.shape[1], self.w2.out_len,
SILU, qtype SILU, qtype

17
python/llm/src/ipex_llm/transformers/models/qwen2.py
View file

@ -310,8 +310,8 @@ def qwen2_attention_forward(
kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx) kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
if should_use_fuse_rope(hidden_states, position_ids, self.training): if should_use_fuse_rope(hidden_states, position_ids, self.training):
import linear_q4_0 import xe_addons
linear_q4_0.rotary_half_inplaced(self.rotary_emb.inv_freq, position_ids, xe_addons.rotary_half_inplaced(self.rotary_emb.inv_freq, position_ids,
query_states, key_states) query_states, key_states)
else: else:
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len) cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
@ -337,18 +337,19 @@ def qwen2_attention_forward(
value_states.to(device, dtype=torch.float16), value_states.to(device, dtype=torch.float16),
is_causal=True).to(hidden_states.dtype) is_causal=True).to(hidden_states.dtype)
elif use_sdp(q_len, kv_seq_len, self.head_dim, query_states): elif use_sdp(q_len, kv_seq_len, self.head_dim, query_states):
import linear_q4_0 import xe_addons
if isinstance(past_key_value, DynamicFp8Cache): if isinstance(past_key_value, DynamicFp8Cache):
attn_output = linear_q4_0.sdp_fp8(query_states, key_states, value_states, attn_output = xe_addons.sdp_fp8(query_states, key_states, value_states,
attention_mask) attention_mask)
else: else:
attn_output = linear_q4_0.sdp(query_states, key_states, value_states, attention_mask) attn_output = xe_addons.sdp(query_states, key_states, value_states,
attention_mask)
elif use_sdp_causal(q_len, kv_seq_len, self.head_dim, query_states, self.training): elif use_sdp_causal(q_len, kv_seq_len, self.head_dim, query_states, self.training):
import linear_q4_0 import xe_addons
if isinstance(past_key_value, DynamicFp8Cache): if isinstance(past_key_value, DynamicFp8Cache):
attn_output = linear_q4_0.sdp_fp8_causal(query_states, key_states, value_states) attn_output = xe_addons.sdp_fp8_causal(query_states, key_states, value_states)
else: else:
attn_output = linear_q4_0.sdp_causal(query_states, key_states, value_states) attn_output = xe_addons.sdp_causal(query_states, key_states, value_states)
else: else:
if isinstance(past_key_value, DynamicFp8Cache): if isinstance(past_key_value, DynamicFp8Cache):
key_states, value_states = restore_fp8_kv_cache(key_states, value_states, key_states, value_states = restore_fp8_kv_cache(key_states, value_states,

20
python/llm/src/ipex_llm/transformers/models/qwen2_moe.py
View file

@ -372,8 +372,8 @@ def qwen2moe_attention_forward_quantized(
self.layer_idx, cache_kwargs) self.layer_idx, cache_kwargs)
if q_len == 1 and query_states.device.type == 'xpu' and not self.training \ if q_len == 1 and query_states.device.type == 'xpu' and not self.training \
and not hidden_states.requires_grad: and not hidden_states.requires_grad:
import linear_q4_0 import xe_addons
attn_weights = linear_q4_0.query_key_fp8_matmul(query_states, key_states) attn_weights = xe_addons.query_key_fp8_matmul(query_states, key_states)
else: else:
key_states, value_states = restore_fp8_kv_cache(key_states, key_states, value_states = restore_fp8_kv_cache(key_states,
value_states, query_states.dtype) value_states, query_states.dtype)
@ -404,8 +404,8 @@ def qwen2moe_attention_forward_quantized(
p=self.attention_dropout, training=self.training) p=self.attention_dropout, training=self.training)
if q_len == 1 and query_states.device.type == 'xpu' and not self.training \ if q_len == 1 and query_states.device.type == 'xpu' and not self.training \
and not hidden_states.requires_grad: and not hidden_states.requires_grad:
import linear_q4_0 import xe_addons
attn_output = linear_q4_0.attn_value_fp8_matmul(attn_weights, value_states) attn_output = xe_addons.attn_value_fp8_matmul(attn_weights, value_states)
else: else:
attn_output = torch.matmul(attn_weights, value_states) attn_output = torch.matmul(attn_weights, value_states)
@ -456,12 +456,12 @@ def qwen2moe_attention_forward_origin(
cache_k = past_key_value.key_cache[self.layer_idx] cache_k = past_key_value.key_cache[self.layer_idx]
cache_v = past_key_value.value_cache[self.layer_idx] cache_v = past_key_value.value_cache[self.layer_idx]
kv_seq_len = cache_k.shape[-2] kv_seq_len = cache_k.shape[-2]
import linear_q4_0 import xe_linear
args = [hidden_states, self.q_proj.weight, self.k_proj.weight, self.v_proj.weight, args = [hidden_states, self.q_proj.weight, self.k_proj.weight, self.v_proj.weight,
self.q_proj.bias, self.k_proj.bias, self.v_proj.bias, position_ids, cache_k, self.q_proj.bias, self.k_proj.bias, self.v_proj.bias, position_ids, cache_k,
cache_v, self.q_proj.weight.qtype, self.v_proj.weight.qtype, kv_seq_len, cache_v, self.q_proj.weight.qtype, self.v_proj.weight.qtype, kv_seq_len,
self.head_dim, self.rotary_emb.base] self.head_dim, self.rotary_emb.base]
query_states, key_states, value_states = linear_q4_0.forward_qkv_bias(*args) query_states, key_states, value_states = xe_linear.forward_qkv_bias(*args)
kv_seq_len += 1 kv_seq_len += 1
if self.layer_idx == 0: if self.layer_idx == 0:
past_key_value._seen_tokens = kv_seq_len past_key_value._seen_tokens = kv_seq_len
@ -613,12 +613,12 @@ def qwen2moe_attention_forward_sdpa(
cache_k = past_key_value.key_cache[self.layer_idx] cache_k = past_key_value.key_cache[self.layer_idx]
cache_v = past_key_value.value_cache[self.layer_idx] cache_v = past_key_value.value_cache[self.layer_idx]
kv_seq_len = cache_k.shape[-2] kv_seq_len = cache_k.shape[-2]
import linear_q4_0 import xe_linear
args = [hidden_states, self.q_proj.weight, self.k_proj.weight, self.v_proj.weight, args = [hidden_states, self.q_proj.weight, self.k_proj.weight, self.v_proj.weight,
self.q_proj.bias, self.k_proj.bias, self.v_proj.bias, position_ids, cache_k, self.q_proj.bias, self.k_proj.bias, self.v_proj.bias, position_ids, cache_k,
cache_v, self.q_proj.weight.qtype, self.v_proj.weight.qtype, kv_seq_len, cache_v, self.q_proj.weight.qtype, self.v_proj.weight.qtype, kv_seq_len,
self.head_dim, self.rotary_emb.base] self.head_dim, self.rotary_emb.base]
query_states, key_states, value_states = linear_q4_0.forward_qkv_bias(*args) query_states, key_states, value_states = xe_linear.forward_qkv_bias(*args)
kv_seq_len += 1 kv_seq_len += 1
if self.layer_idx == 0: if self.layer_idx == 0:
past_key_value._seen_tokens = kv_seq_len past_key_value._seen_tokens = kv_seq_len
@ -765,8 +765,8 @@ def qwen2moe_moeblock_forward(self, hidden_states: torch.Tensor):
elif bs < 256 and hidden_states.device.type == 'xpu': elif bs < 256 and hidden_states.device.type == 'xpu':
final_hidden_states = torch.zeros((batch_size * sequence_length, hidden_dim), final_hidden_states = torch.zeros((batch_size * sequence_length, hidden_dim),
dtype=hidden_states.dtype, device=hidden_states.device) dtype=hidden_states.dtype, device=hidden_states.device)
import linear_q4_0 import xe_linear
indexes = linear_q4_0.get_moe_indexes(selected_experts.to(torch.int32).cpu(), 60) indexes = xe_linear.get_moe_indexes(selected_experts.to(torch.int32).cpu(), 60)
for expert_idx in range(self.num_experts): for expert_idx in range(self.num_experts):
expert_layer = self.experts[expert_idx] expert_layer = self.experts[expert_idx]
idx_list = indexes[0][expert_idx] idx_list = indexes[0][expert_idx]

4
python/llm/src/ipex_llm/transformers/models/qwen_vl.py
View file

@ -162,8 +162,8 @@ def qwen_attention_forward_vl(
if not self.training and not hidden_states.requires_grad and \ if not self.training and not hidden_states.requires_grad and \
use_sdp(q_len, key.shape[2], self.head_dim, query): use_sdp(q_len, key.shape[2], self.head_dim, query):
import linear_q4_0 import xe_addons
attn_output = linear_q4_0.sdp(query, key, value, attention_mask) attn_output = xe_addons.sdp(query, key, value, attention_mask)
attn_output = attn_output.view(query.shape) attn_output = attn_output.view(query.shape)
attn_output = attn_output.transpose(1, 2) attn_output = attn_output.transpose(1, 2)
attn_weight = None attn_weight = None

12
python/llm/src/ipex_llm/transformers/models/rwkv4.py
View file

@ -56,8 +56,8 @@ def extract_key_value(self, hidden, state=None):
self.time_mix_receptance.data, self.time_mix_receptance.data,
]).to(dtype=hidden.dtype) ]).to(dtype=hidden.dtype)
import linear_q4_0 import xe_linear
mixed_result = linear_q4_0.rwkv_time_shift(hidden, shifted, self.mixed_mix) mixed_result = xe_linear.rwkv_time_shift(hidden, shifted, self.mixed_mix)
key, value, receptance = mixed_result key, value, receptance = mixed_result
key = self.key(key) key = self.key(key)
@ -92,8 +92,8 @@ def rwkv_linear_attention_xpu(
time_decay = -torch.exp(time_decay) time_decay = -torch.exp(time_decay)
# `num_state`, `den_state`, `max_state` will be modified during this call # `num_state`, `den_state`, `max_state` will be modified during this call
import linear_q4_0 import xe_linear
output = linear_q4_0.rwkv_linear_attention_v4( output = xe_linear.rwkv_linear_attention_v4(
time_decay, time_decay,
time_first, time_first,
key, key,
@ -167,8 +167,8 @@ def rwkv_ffn_forward(
self.mixed_mix = torch.cat([self.time_mix_key.data, self.mixed_mix = torch.cat([self.time_mix_key.data,
self.time_mix_receptance.data]).to(dtype=hidden.dtype) self.time_mix_receptance.data]).to(dtype=hidden.dtype)
import linear_q4_0 import xe_linear
mixed_result = linear_q4_0.rwkv_time_shift(hidden, shifted, self.mixed_mix) mixed_result = xe_linear.rwkv_time_shift(hidden, shifted, self.mixed_mix)
key, receptance = mixed_result key, receptance = mixed_result
key = torch.square(torch.relu(self.key(key))) key = torch.square(torch.relu(self.key(key)))

12
python/llm/src/ipex_llm/transformers/models/rwkv5.py
View file

@ -58,8 +58,8 @@ def extract_key_value(self, hidden, state=None):
self.time_mix_gate.data, self.time_mix_gate.data,
]).to(dtype=hidden.dtype) ]).to(dtype=hidden.dtype)
import linear_q4_0 import xe_linear
mixed_result = linear_q4_0.rwkv_time_shift(hidden, shifted, self.mixed_mix) mixed_result = xe_linear.rwkv_time_shift(hidden, shifted, self.mixed_mix)
key, value, receptance, gate = mixed_result key, value, receptance, gate = mixed_result
key = self.key(key) key = self.key(key)
@ -98,8 +98,8 @@ def rwkv_linear_attention_xpu(
time_first = time_first.float() time_first = time_first.float()
# `state` will be updated inplaced during this call # `state` will be updated inplaced during this call
import linear_q4_0 import xe_linear
out = linear_q4_0.rwkv_linear_attention_v5( out = xe_linear.rwkv_linear_attention_v5(
time_decay, time_decay,
time_first, time_first,
receptance, receptance,
@ -236,8 +236,8 @@ def rwkv_ffn_forward_wrapper(origin_rwkv_ffn_forward):
self.mixed_mix = torch.cat([self.time_mix_key.data, self.mixed_mix = torch.cat([self.time_mix_key.data,
self.time_mix_receptance.data]).to(dtype=hidden.dtype) self.time_mix_receptance.data]).to(dtype=hidden.dtype)
import linear_q4_0 import xe_linear
mixed_result = linear_q4_0.rwkv_time_shift(hidden, shifted, self.mixed_mix) mixed_result = xe_linear.rwkv_time_shift(hidden, shifted, self.mixed_mix)
key, receptance = mixed_result key, receptance = mixed_result
key = torch.square(torch.relu(self.key(key))) key = torch.square(torch.relu(self.key(key)))

13
python/llm/src/ipex_llm/transformers/models/stablelm.py
View file

@ -267,8 +267,9 @@ def stablelm_attention_forward_original(
attn_weights = None attn_weights = None
elif not self.training and not hidden_states.requires_grad and \ elif not self.training and not hidden_states.requires_grad and \
use_sdp(q_len, key_states.shape[2], self.head_dim, query_states): use_sdp(q_len, key_states.shape[2], self.head_dim, query_states):
import linear_q4_0 import xe_addons
attn_output = linear_q4_0.sdp(query_states, key_states, value_states, attention_mask) attn_output = xe_addons.sdp(query_states, key_states, value_states,
attention_mask)
attn_output = attn_output.view(query_states.shape) attn_output = attn_output.view(query_states.shape)
attn_weights = None attn_weights = None
else: else:
@ -420,8 +421,8 @@ def stablelm_attention_forward_quantized(
value_states = repeat_kv(value_states, self.num_key_value_groups) value_states = repeat_kv(value_states, self.num_key_value_groups)
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) attn_weights = torch.matmul(query_states, key_states.transpose(2, 3))
else: else:
import linear_q4_0 import xe_addons
attn_weights = linear_q4_0.query_key_fp8_matmul(query_states, key_states) attn_weights = xe_addons.query_key_fp8_matmul(query_states, key_states)
attn_weights = attn_weights / math.sqrt(self.head_dim) attn_weights = attn_weights / math.sqrt(self.head_dim)
@ -444,8 +445,8 @@ def stablelm_attention_forward_quantized(
if query_states.size(2) != 1 or query_states.device.type != 'xpu': if query_states.size(2) != 1 or query_states.device.type != 'xpu':
attn_output = torch.matmul(attn_weights, value_states) attn_output = torch.matmul(attn_weights, value_states)
else: else:
import linear_q4_0 import xe_addons
attn_output = linear_q4_0.attn_value_fp8_matmul(attn_weights, value_states) attn_output = xe_addons.attn_value_fp8_matmul(attn_weights, value_states)
attn_output_size = (bsz, self.num_heads, q_len, self.head_dim) attn_output_size = (bsz, self.num_heads, q_len, self.head_dim)
invalidInputError(attn_output.size() == attn_output_size, invalidInputError(attn_output.size() == attn_output_size,

5
python/llm/src/ipex_llm/transformers/models/starcoder2.py
View file

@ -135,8 +135,9 @@ def attention_forward(
self.layer_idx, None) self.layer_idx, None)
if use_quantize_kv and q_len == 1: if use_quantize_kv and q_len == 1:
import linear_q4_0 import xe_addons
attn_output = linear_q4_0.sdp_fp8(query_states, key_states, value_states, attention_mask) attn_output = xe_addons.sdp_fp8(query_states, key_states, value_states,
attention_mask)
attn_weights = None attn_weights = None
else: else:
if use_quantize_kv: if use_quantize_kv:

26
python/llm/src/ipex_llm/transformers/models/utils.py
View file

@ -20,7 +20,8 @@ import warnings
from ipex_llm.utils.common import invalidInputError from ipex_llm.utils.common import invalidInputError
from ipex_llm.ggml.quantize import ggml_tensor_qtype from ipex_llm.ggml.quantize import ggml_tensor_qtype
from ipex_llm.transformers.utils import get_ipex_version, get_xpu_device_type from ipex_llm.transformers.utils import get_ipex_version, get_xpu_device_type
from ipex_llm.transformers.low_bit_linear import SYM_INT4, SYM_INT8, FP8E5, IQ2_XXS, FP4, FP8E4, FP6 from ipex_llm.transformers.low_bit_linear import SYM_INT4, SYM_INT8, FP8E5, IQ2_XXS, FP4, FP8E4,\
FP6, ASYM_INT4
from ipex_llm.transformers.convert import is_deepspeed_available from ipex_llm.transformers.convert import is_deepspeed_available
FP8_KV_ALLOC_LENGTH = 512 FP8_KV_ALLOC_LENGTH = 512
@ -128,8 +129,8 @@ def append_fp8_kv_cache(k_cache, v_cache, key, value):
new_k_cache = k_cache.as_strided(new_size, k_cache.stride(), storage_offset=0) new_k_cache = k_cache.as_strided(new_size, k_cache.stride(), storage_offset=0)
new_v_cache = v_cache.as_strided(new_size, v_cache.stride(), storage_offset=0) new_v_cache = v_cache.as_strided(new_size, v_cache.stride(), storage_offset=0)
import linear_q4_0 import xe_addons
linear_q4_0.quantize_key_value(key, value, xe_addons.quantize_key_value(key, value,
new_k_cache[:, :, cur_length:new_length, :], new_k_cache[:, :, cur_length:new_length, :],
new_v_cache[:, :, cur_length:new_length, :]) new_v_cache[:, :, cur_length:new_length, :])
@ -140,8 +141,8 @@ def restore_fp8_kv_cache(k_cache, v_cache, dtype):
key_states = torch.empty(k_cache.shape, device=k_cache.device, dtype=dtype) key_states = torch.empty(k_cache.shape, device=k_cache.device, dtype=dtype)
value_states = torch.empty(v_cache.shape, device=v_cache.device, dtype=dtype) value_states = torch.empty(v_cache.shape, device=v_cache.device, dtype=dtype)
import linear_q4_0 import xe_addons
linear_q4_0.dequantize_key_value(k_cache, v_cache, key_states, value_states) xe_addons.dequantize_key_value(k_cache, v_cache, key_states, value_states)
return key_states, value_states return key_states, value_states
@ -211,12 +212,12 @@ def apply_rotary_pos_emb_no_cache_xpu(q, k, position_ids, model_family, rope_the
if q.device.type != "xpu": if q.device.type != "xpu":
invalidInputError(False, invalidInputError(False,
f"only xpu is supported in this function") f"only xpu is supported in this function")
import linear_q4_0 import xe_addons
q_embed = torch.empty(q.shape, dtype=q.dtype, device=q.device) q_embed = torch.empty(q.shape, dtype=q.dtype, device=q.device)
k_embed = torch.empty(k.shape, dtype=k.dtype, device=k.device) k_embed = torch.empty(k.shape, dtype=k.dtype, device=k.device)
if model_family in ["llama", "baichuan", "internlm", "aquila", "gpt_neox", "mistral", if model_family in ["llama", "baichuan", "internlm", "aquila", "gpt_neox", "mistral",
"mixtral"]: "mixtral"]:
linear_q4_0.apply_rotary_embedding_half_q_and_k(q, k, position_ids, xe_addons.apply_rotary_embedding_half_q_and_k(q, k, position_ids,
q_embed, k_embed, rope_theta) q_embed, k_embed, rope_theta)
return q_embed, k_embed return q_embed, k_embed
else: else:
@ -228,11 +229,12 @@ def apply_rotary_pos_emb_cache_freq_xpu(q, k, sin, cos, model_family, position_i
if q.device.type != "xpu": if q.device.type != "xpu":
invalidInputError(False, invalidInputError(False,
f"only xpu is supported in this function") f"only xpu is supported in this function")
import linear_q4_0 import xe_addons
q_embed = torch.empty(q.shape, dtype=q.dtype, device=q.device) q_embed = torch.empty(q.shape, dtype=q.dtype, device=q.device)
k_embed = torch.empty(k.shape, dtype=k.dtype, device=k.device) k_embed = torch.empty(k.shape, dtype=k.dtype, device=k.device)
if model_family in ["qwen", "mixtral"]: if model_family in ["qwen", "mixtral"]:
linear_q4_0.apply_rotary_embedding_half_q_and_k_cache_freq(q, k, sin, cos, q_embed, k_embed) xe_addons.apply_rotary_embedding_half_q_and_k_cache_freq(q, k, sin, cos,
q_embed, k_embed)
elif model_family in ["qwen2", "yuan", "stablelm", "qwen2_moe", "internlm"]: elif model_family in ["qwen2", "yuan", "stablelm", "qwen2_moe", "internlm"]:
cos = cos.to(q.dtype) cos = cos.to(q.dtype)
sin = sin.to(q.dtype) sin = sin.to(q.dtype)
@ -240,11 +242,13 @@ def apply_rotary_pos_emb_cache_freq_xpu(q, k, sin, cos, model_family, position_i
sin = sin.squeeze(1).squeeze(0) # [seq_len, dim] sin = sin.squeeze(1).squeeze(0) # [seq_len, dim]
cos = cos[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim] cos = cos[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
sin = sin[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim] sin = sin[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
linear_q4_0.apply_rotary_embedding_half_q_and_k_cache_freq(q, k, sin, cos, q_embed, k_embed) xe_addons.apply_rotary_embedding_half_q_and_k_cache_freq(q, k, sin, cos,
q_embed, k_embed)
elif model_family in ["gemma", "phi3"]: elif model_family in ["gemma", "phi3"]:
cos = cos.unsqueeze(1) cos = cos.unsqueeze(1)
sin = sin.unsqueeze(1) sin = sin.unsqueeze(1)
linear_q4_0.apply_rotary_embedding_half_q_and_k_cache_freq(q, k, sin, cos, q_embed, k_embed) xe_addons.apply_rotary_embedding_half_q_and_k_cache_freq(q, k, sin, cos,
q_embed, k_embed)
else: else:
invalidInputError(False, invalidInputError(False,
f"{model_family} is not supported.") f"{model_family} is not supported.")

12
python/llm/src/ipex_llm/transformers/models/yuan.py
View file

@ -97,10 +97,10 @@ def yuan_mlp_forward(
bsz, hidden_size = x_2d.shape bsz, hidden_size = x_2d.shape
qtype = getattr(self.up_proj, "qtype", None) qtype = getattr(self.up_proj, "qtype", None)
if mlp_fusion_check(x_2d, qtype, self.training): if mlp_fusion_check(x_2d, qtype, self.training):
import linear_q4_0 import xe_linear
if not x_2d.is_contiguous(): if not x_2d.is_contiguous():
x_2d = x_2d.contiguous() x_2d = x_2d.contiguous()
out = self.down_proj(linear_q4_0.mlp_forward_xpu( out = self.down_proj(xe_linear.mlp_forward_xpu(
x_2d, self.up_proj.weight.data, self.gate_proj.weight.data, x_2d, self.up_proj.weight.data, self.gate_proj.weight.data,
x_2d.shape[0], x_2d.shape[1], self.up_proj.out_len, x_2d.shape[0], x_2d.shape[1], self.up_proj.out_len,
SILU, qtype SILU, qtype
@ -268,8 +268,8 @@ def yuan_attention_forward_quantized(
query_states.dtype) query_states.dtype)
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) attn_weights = torch.matmul(query_states, key_states.transpose(2, 3))
else: else:
import linear_q4_0 import xe_addons
attn_weights = linear_q4_0.query_key_fp8_matmul(query_states, key_states) attn_weights = xe_addons.query_key_fp8_matmul(query_states, key_states)
attn_weights = attn_weights / math.sqrt(self.head_dim) attn_weights = attn_weights / math.sqrt(self.head_dim)
@ -292,8 +292,8 @@ def yuan_attention_forward_quantized(
if query_states.size(2) != 1 or device.type != 'xpu': if query_states.size(2) != 1 or device.type != 'xpu':
attn_output = torch.matmul(attn_weights, value_states) attn_output = torch.matmul(attn_weights, value_states)
else: else:
import linear_q4_0 import xe_addons
attn_output = linear_q4_0.attn_value_fp8_matmul(attn_weights, value_states) attn_output = xe_addons.attn_value_fp8_matmul(attn_weights, value_states)
invalidInputError(attn_output.size() == (bsz, self.num_heads, q_len, self.head_dim), invalidInputError(attn_output.size() == (bsz, self.num_heads, q_len, self.head_dim),
"`attn_output` should be of size " "`attn_output` should be of size "