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

refactor chatglm2/3 (#11290)

Browse source
This commit is contained in:
Yishuo Wang 2024-06-13 12:22:58 +08:00 committed by GitHub
parent ea372cc472
commit 01fe0fc1a2
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 152 additions and 509 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

651
python/llm/src/ipex_llm/transformers/models/chatglm2.py
View file

@ -19,136 +19,26 @@
import math
import torch
from typing import Optional, Tuple, List
import torch.nn.functional as F
from typing import Optional, Tuple
from transformers.modeling_outputs import BaseModelOutputWithPast
from ipex_llm.transformers.models.utils import init_kv_cache, extend_kv_cache, append_kv_cache
from ipex_llm.transformers.models.utils import init_fp8_kv_cache, append_fp8_kv_cache, \
restore_fp8_kv_cache, use_quantize_kv_cache, use_flash_attention
from ipex_llm.transformers.models.utils import use_sdp
from ipex_llm.utils.common.log4Error import invalidInputError
from ipex_llm.transformers.models.utils import restore_fp8_kv_cache, update_past_key_value
from ipex_llm.transformers.models.utils import use_quantize_kv_cache, use_sdp, use_sdp_causal
from ipex_llm.transformers.models.utils import should_use_fuse_rope, apply_rotary_pos_emb
import os
KV_CACHE_ALLOC_BLOCK_LENGTH = int(os.environ.get("KV_CACHE_ALLOC_BLOCK_LENGTH", 256))
KV_CACHE_ALLOC_MIN_LENGTH = 512
def split_tensor_along_last_dim(
tensor: torch.Tensor,
num_partitions: int,
contiguous_split_chunks: bool = False,
) -> List[torch.Tensor]:
"""Split a tensor along its last dimension.
Arguments:
tensor: input tensor.
num_partitions: number of partitions to split the tensor
contiguous_split_chunks: If True, make each chunk contiguous
in memory.
Returns:
A list of Tensors
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
"""
# Get the size and dimension.
last_dim = tensor.dim() - 1
last_dim_size = tensor.size()[last_dim] // num_partitions
# Split.
tensor_list = torch.split(tensor, last_dim_size, dim=last_dim)
# Note: torch.split does not create contiguous tensors by default.
if contiguous_split_chunks:
return tuple(chunk.contiguous() for chunk in tensor_list)
return tensor_list
def glm_sdpa(query, key, value, attention_mask=None, is_causal=False):
if use_flash_attention(query, key, attention_mask) or query.device.type == 'cpu':
context_layer = F.scaled_dot_product_attention(query.to(key.dtype),
key,
value,
attention_mask,
is_causal=is_causal).to(key.dtype)
else:
# attention_mask is not None only when past_key_value is not None and q_len > 1
if attention_mask is not None:
attn_bias = torch.zeros(attention_mask.shape, dtype=query.dtype,
device=query.device)
attention_mask = ~attention_mask
if attention_mask.dtype == torch.bool:
attn_bias.masked_fill_(attention_mask.logical_not(), float("-inf"))
else:
attn_bias += attention_mask
elif is_causal:
L, S = query.size(-2), key.size(-2)
attn_bias = torch.zeros(L, S, dtype=query.dtype, device=query.device)
temp_mask = torch.ones(L, S, dtype=torch.bool, device=query.device).tril(diagonal=0)
attn_bias.masked_fill_(temp_mask.logical_not(), float("-inf"))
attn_bias.to(key.dtype)
else:
attn_bias = None
if use_sdp(query.shape[2], key.shape[2],
query.shape[-1], query):
import xe_addons
attn_output = xe_addons.sdp(query, key, value, attn_bias)
context_layer = attn_output.view(query.shape)
else:
head_dim = query.size(-1)
attn = torch.matmul(query.to(key.dtype),
key.transpose(2, 3)) / math.sqrt(head_dim)
if attn_bias is not None:
attn += attn_bias
attn = F.softmax(attn, dim=-1,
dtype=torch.float32).to(value.dtype)
context_layer = torch.matmul(attn, value)
return context_layer
@torch.jit.script
def apply_rotary_pos_emb_chatglm(x: torch.Tensor, rope_cache: torch.Tensor) -> torch.Tensor:
# x: [sq, b, np, hn]
sq, b, np, hn = x.size(0), x.size(1), x.size(2), x.size(3)
rot_dim = rope_cache.shape[-2] * 2
x, x_pass = x[..., :rot_dim], x[..., rot_dim:]
# truncate to support variable sizes
rope_cache = rope_cache[:sq]
xshaped = x.reshape(sq, -1, np, rot_dim // 2, 2)
rope_cache = rope_cache.view(sq, -1, 1, xshaped.size(3), 2)
x_out2 = torch.stack(
[
xshaped[..., 0] * rope_cache[..., 0] - xshaped[..., 1] * rope_cache[..., 1],
xshaped[..., 1] * rope_cache[..., 0] + xshaped[..., 0] * rope_cache[..., 1],
],
-1,
)
x_out2 = x_out2.flatten(3)
return torch.cat((x_out2, x_pass), dim=-1)
def repeat_kv(key: torch.Tensor, value: torch.Tensor, n_head: int) -> (torch.Tensor, torch.Tensor):
# key, value's shape: [bs, n_kv_head, seq_len, head_dim] -> [bs, n_head, seq_len, head_dim]
batch_size, n_kv_head, seq_len, head_dim = key.shape
key = key.unsqueeze(2)
key = key.expand(-1, -1, n_head // n_kv_head, -1, -1)
key = key.contiguous().view(batch_size, n_head, seq_len, head_dim)
value = value.unsqueeze(2)
value = value.expand(-1, -1, n_head // n_kv_head, -1, -1)
value = value.contiguous().view(batch_size, n_head, seq_len, head_dim)
return key, value
def should_split_qkv_tensor(query_layer, bsz, n_head, seq_len):
if os.environ.get("IPEX_LLM_SPLIT_QKV", None) is not None:
return os.environ.get("IPEX_LLM_SPLIT_QKV", None) == "1"
elif query_layer.dtype == torch.float16 and query_layer.shape[2] >= 5000:
# split tensor for memory block limitation
# support fp16 and set input length threshold at 5000 for now
return True
elif query_layer.element_size()*bsz*n_head*seq_len*seq_len >= 4*1024**3:
# attn_weight size larger than memory block limitation 4GB
return True
return False
This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states
go from (batch, num_key_value_heads, seqlen, head_dim) to
(batch, num_attention_heads, seqlen, head_dim)
"""
batch, num_key_value_heads, slen, head_dim = hidden_states.shape
if n_rep == 1:
return hidden_states
hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads,
n_rep, slen, head_dim)
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
def chatglm_rms_norm_forward(self, hidden_states):
@ -166,16 +56,16 @@ def chatglm_rms_norm_forward(self, hidden_states):
def chatglm2_model_forward(
self,
input_ids,
position_ids: Optional[torch.Tensor]=None,
attention_mask: Optional[torch.BoolTensor]=None,
full_attention_mask: Optional[torch.BoolTensor]=None,
past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]]=None,
inputs_embeds: Optional[torch.Tensor]=None,
use_cache: Optional[bool]=None,
output_hidden_states: Optional[bool]=None,
return_dict: Optional[bool]=None,
self,
input_ids,
position_ids: Optional[torch.Tensor]=None,
attention_mask: Optional[torch.BoolTensor]=None,
full_attention_mask: Optional[torch.BoolTensor]=None,
past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]]=None,
inputs_embeds: Optional[torch.Tensor]=None,
use_cache: Optional[bool]=None,
output_hidden_states: Optional[bool]=None,
return_dict: Optional[bool]=None,
):
output_hidden_states = (
output_hidden_states if output_hidden_states is not None
@ -196,33 +86,51 @@ def chatglm2_model_forward(
past_key_values,
padding_mask=attention_mask)
use_fuse_rope = input_ids.device.type == "xpu"
use_fuse_rope = use_fuse_rope and not self.training
# ipex-llm changes begin
# 1. replace `rotary_pos_emb` with `inv_freq` and `position_ids`
# 2. generate `causal_mask` and replace `full_attention_mask` with it
if position_ids is None:
if past_key_values is None:
position_ids = torch.arange(seq_length, dtype=torch.int64, device=inputs_embeds.device)
else:
kv_length = past_key_values[0][0].size(0)
position_ids = torch.arange(kv_length, kv_length + seq_length,
dtype=torch.int64, device=inputs_embeds.device)
position_ids = position_ids.repeat(batch_size, 1)
# Rotary positional embeddings
rotary_pos_emb = self.rotary_pos_emb(self.seq_length)
if position_ids is not None:
rotary_pos_emb = rotary_pos_emb[position_ids]
if getattr(self.rotary_pos_emb, "cached_dtype", None) != inputs_embeds.dtype:
rot_dim = self.rotary_pos_emb.dim
inv_freq = 1.0 / (10000 ** (torch.arange(0, rot_dim, 2,
device=inputs_embeds.device,
dtype=inputs_embeds.dtype) / rot_dim))
self.rotary_pos_emb.register_buffer("inv_freq", inv_freq, persistent=False)
self.rotary_pos_emb.cached_dtype = inputs_embeds.dtype
# `full_attention_mask` is not None only when
# `past_key_values` is not None and `seq_length` > 1
if full_attention_mask is not None:
causal_mask = torch.zeros([batch_size, 1, seq_length, full_attention_mask.size(-1)],
dtype=inputs_embeds.dtype, device=inputs_embeds.device)
mask_value = torch.finfo(inputs_embeds.dtype).min
causal_mask.masked_fill_(full_attention_mask, mask_value)
elif self.training or (inputs_embeds.device.type != "xpu" and past_key_values is None):
full_attention_mask = self.get_masks(input_ids,
past_key_values,
padding_mask=attention_mask)
causal_mask = torch.zeros([batch_size, 1, seq_length, full_attention_mask.size(-1)],
dtype=inputs_embeds.dtype, device=inputs_embeds.device)
mask_value = torch.finfo(inputs_embeds.dtype).min
causal_mask.masked_fill_(full_attention_mask, mask_value)
else:
rotary_pos_emb = rotary_pos_emb[None, :seq_length]
if use_fuse_rope:
# Repeat cos sin here, call only once for each token.
# Chatglm2's rotary embedding is similar to gptj's, is rotate_every_two.
# If put this to attension forward, it will generate too many times.
cos, sin = rotary_pos_emb.split(rotary_pos_emb.shape[-1] // 2, dim=-1)
cos = cos.squeeze(-1)
sin = sin.squeeze(-1)
cos = torch.repeat_interleave(cos[:, :, None, :], 2, 3)
sin = torch.repeat_interleave(sin[:, :, None, :], 2, 3)
rotary_pos_emb = (cos, sin)
else:
rotary_pos_emb = rotary_pos_emb.transpose(0, 1).contiguous()
causal_mask = None
# Run encoder.
hidden_states, presents, all_hidden_states, all_self_attentions = self.encoder(
inputs_embeds, full_attention_mask, rotary_pos_emb=rotary_pos_emb,
inputs_embeds, causal_mask,
rotary_pos_emb=(self.rotary_pos_emb.inv_freq, position_ids),
kv_caches=past_key_values, use_cache=use_cache, output_hidden_states=output_hidden_states
)
# ipex-llm changes end
if not return_dict:
return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions]
@ -239,364 +147,105 @@ def chatglm2_model_forward(
def chatglm2_attention_forward(
self, hidden_states, attention_mask, rotary_pos_emb, kv_cache=None, use_cache=True
):
if use_quantize_kv_cache(self.query_key_value, hidden_states.transpose(0, 1)):
forward_function = chatglm2_quantized_attention_forward_8eb45c
else:
forward_function = chatglm2_attention_forward_8eb45c
return forward_function(
self=self,
hidden_states=hidden_states,
attention_mask=attention_mask,
rotary_pos_emb=rotary_pos_emb,
kv_cache=kv_cache,
use_cache=use_cache
)
# hidden_states: [seq_len, bsz, head_dim]
q_len, bsz, _ = hidden_states.size()
def chatglm2_quantized_attention_forward_8eb45c(
self, hidden_states, attention_mask, rotary_pos_emb, kv_cache=None, use_cache=True
):
# hidden_states: [seq_len, bs, head_dim]
mixed_x_layer = self.query_key_value(hidden_states)
# kv_cache: [seq_len, bsz, n_kv_head, head_dim] ->
# past_key_value: [bsz, n_kv_head, seq_len, head_dim]
past_key_value = None if kv_cache is None else (kv_cache[0].permute(1, 2, 0, 3),
kv_cache[1].permute(1, 2, 0, 3))
n_head = self.num_attention_heads_per_partition
n_kv_head = self.num_multi_query_groups_per_partition if self.multi_query_attention else n_head
head_dim = self.hidden_size_per_attention_head
query_layer, key_layer, value_layer = mixed_x_layer.split(
[n_head * head_dim, n_kv_head * head_dim, n_kv_head * head_dim],
dim=-1,
qkv = self.query_key_value(hidden_states)
qkv = qkv.view(q_len, bsz, n_head + 2 * n_kv_head, head_dim)
# [seq_len, bsz, n_head, head_dim] -> [bsz, n_head, seq_len, head_dim]
qkv = qkv.permute(1, 2, 0, 3)
query_states, key_states, value_states = qkv.split([n_head,
n_kv_head,
n_kv_head], dim=1)
kv_seq_len = key_states.shape[2]
if past_key_value is not None:
kv_seq_len += past_key_value[0].shape[2]
# IPEX-LLM OPT: fuse rope
inv_freq, position_ids = rotary_pos_emb
rot_dim = inv_freq.size(-1) * 2
if should_use_fuse_rope(hidden_states, rotary_pos_emb[1], self.training):
import xe_addons
xe_addons.rotary_two_inplaced(inv_freq, position_ids,
query_states[..., :rot_dim], key_states[..., :rot_dim])
else:
idx_theta = torch.outer(position_ids[0].float(),
inv_freq.float()).to(hidden_states.dtype)
idx_theta = idx_theta.unsqueeze(0).unsqueeze(0)
cos = torch.cos(idx_theta).repeat_interleave(2, -1)
sin = torch.sin(idx_theta).repeat_interleave(2, -1)
q_rot, k_rot = apply_rotary_pos_emb(query_states[..., :rot_dim], key_states[..., :rot_dim],
cos, sin, position_ids, "chatglm")
query_states[..., :rot_dim] = q_rot[...]
key_states[..., :rot_dim] = k_rot[...]
# IPEX-LLM OPT: kv cache and quantize kv
use_quantize_kv = use_quantize_kv_cache(self.query_key_value, hidden_states)
key_states, value_states = update_past_key_value(
past_key_value, key_states, value_states,
kv_seq_len, use_quantize_kv, hidden_states.device
)
query_layer = query_layer.view(query_layer.shape[:-1] + (n_head, head_dim))
key_layer = key_layer.view(key_layer.shape[:-1] + (n_kv_head, head_dim))
value_layer = value_layer.view(value_layer.shape[:-1] + (n_kv_head, head_dim))
# query, key, value's shape: [seq_len, bs, n_head/n_kv_head, head_dim]
# past_key_value: [bsz, n_kv_head, seq_len, head_dim] -> [seq_len, bsz, n_kv_head, head_dim]
past_key_value = (key_states.permute(2, 0, 1, 3),
value_states.permute(2, 0, 1, 3)) if use_cache else None
# apply relative positional encoding (rotary embedding)
if rotary_pos_emb is not None:
if len(rotary_pos_emb) == 2 and isinstance(rotary_pos_emb, tuple):
# use_fuse_rope, see chatglm2_model_forward
cos, sin = rotary_pos_emb
rot_dim = cos.shape[-1]
query_layer = query_layer.transpose(0, 1)
key_layer = key_layer.transpose(0, 1)
query_layer_cur = query_layer[..., :rot_dim]
key_layer_cur = key_layer[..., :rot_dim]
# ipex_llm's apply_rotary_embedding can change the origin storage,
# so query_layer will get the result directly.
torch.ops.torch_ipex.apply_rotary_embedding(query_layer_cur, sin, cos, query_layer_cur)
torch.ops.torch_ipex.apply_rotary_embedding(key_layer_cur, sin, cos, key_layer_cur)
query_layer = query_layer.transpose(0, 1)
key_layer = key_layer.transpose(0, 1)
# IPEX-LLM OPT: sdp
attn_weights = None
if use_sdp(q_len, kv_seq_len, head_dim, query_states):
import xe_addons
if use_quantize_kv:
attn_output = xe_addons.sdp_fp8(query_states, key_states, value_states, attention_mask)
else:
query_layer = apply_rotary_pos_emb_chatglm(query_layer, rotary_pos_emb)
key_layer = apply_rotary_pos_emb_chatglm(key_layer, rotary_pos_emb)
query_layer = query_layer.permute(1, 2, 0, 3)
key_layer = key_layer.permute(1, 2, 0, 3)
value_layer = value_layer.permute(1, 2, 0, 3)
# query, key, value's shape: [bs, n_head/n_kv_head, seq_len, head_dim]
batch_size, _, seq_len, _ = query_layer.shape
if kv_cache is None:
# first token
if self.multi_query_attention:
key, value = repeat_kv(key_layer, value_layer, n_head)
attn_output = xe_addons.sdp(query_states, key_states, value_states, attention_mask)
elif use_sdp_causal(q_len, kv_seq_len, head_dim, query_states, self.training):
import xe_addons
if use_quantize_kv:
attn_output = xe_addons.sdp_fp8_causal(query_states, key_states, value_states,
attention_mask)
else:
key, value = key_layer, value_layer
if should_split_qkv_tensor(query_layer, batch_size, n_head, seq_len):
# split second dim to block size = 8
block_size = 8
query_split = torch.split(query_layer, block_size, dim=1)
key_split = torch.split(key, block_size, dim=1)
value_split = torch.split(value, block_size, dim=1)
results = []
for q, k, v in zip(query_split, key_split, value_split):
result = glm_sdpa(q, k, v, is_causal=True)
results.append(result)
context_layer = torch.cat(results, dim=1)
attn_output = xe_addons.sdp_causal(query_states, key_states, value_states,
attention_mask)
elif query_states.device.type == "cpu":
# repeat k/v heads if n_kv_heads < n_heads
key_states = repeat_kv(key_states, n_head // n_kv_head)
value_states = repeat_kv(value_states, n_head // n_kv_head)
if q_len == kv_seq_len:
attn_output = torch.nn.functional.scaled_dot_product_attention(
query_states, key_states, value_states, is_causal=True
)
else:
context_layer = glm_sdpa(query_layer, key, value, is_causal=True)
context_layer = context_layer.to(query_layer.dtype)
if use_cache:
k_cache, v_cache = init_fp8_kv_cache(batch_size,
n_kv_head,
seq_len,
head_dim,
query_layer.device)
k_cache, v_cache = append_fp8_kv_cache(k_cache, v_cache, key_layer, value_layer)
attn_output = torch.nn.functional.scaled_dot_product_attention(
query_states, key_states, value_states, attention_mask
)
else:
k_cache, v_cache = kv_cache
k_cache = k_cache.permute(1, 2, 0, 3)
v_cache = v_cache.permute(1, 2, 0, 3)
# k_cache, v_cache's shape: [bs, n_kv_head, seq_len, head_dim]
k_cache, v_cache = append_fp8_kv_cache(k_cache, v_cache, key_layer, value_layer)
if use_quantize_kv:
key_states, value_states = restore_fp8_kv_cache(key_states, value_states,
query_states.dtype)
# repeat k/v heads if n_kv_heads < n_heads
key_states = repeat_kv(key_states, n_head // n_kv_head)
value_states = repeat_kv(value_states, n_head // n_kv_head)
attn_weights = torch.matmul(query_states,
key_states.transpose(2, 3)) / math.sqrt(head_dim)
if attention_mask is not None:
attention_mask = ~attention_mask
attn_bias = torch.zeros(attention_mask.shape, dtype=query_layer.dtype,
device=query_layer.device)
if attention_mask.dtype == torch.bool:
attn_bias.masked_fill_(attention_mask.logical_not(), float("-inf"))
else:
attn_bias += attention_mask
else:
attn_bias = None
attn_weights = attn_weights + attention_mask
attn_weights = torch.nn.functional.softmax(attn_weights, dim=-1,
dtype=torch.float32).to(value_states.dtype)
attn_output = torch.matmul(attn_weights, value_states)
if seq_len != 1:
key, value = restore_fp8_kv_cache(k_cache, v_cache, query_layer.dtype)
key, value = repeat_kv(key, value, n_head)
attn = torch.matmul(query_layer, key.transpose(2, 3)) / math.sqrt(head_dim)
if attn_bias is not None:
attn += attn_bias
attn = F.softmax(attn, dim=-1, dtype=torch.float32)
context_layer = torch.matmul(attn.to(value.dtype), value)
else:
key, value = k_cache, v_cache
import xe_addons
context_layer = xe_addons.sdp_fp8(query_layer, key, value, attn_bias)
# context_layer's shape: [bsz, n_head, seq_len, head_dim] -> [seq_len, bsz, n_head * head_dim]
attn_output = attn_output.permute(2, 0, 1, 3).contiguous().view(q_len, bsz, n_head * head_dim)
output = self.dense(attn_output)
# 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)
if use_cache:
kv_cache = (k_cache.permute(2, 0, 1, 3), v_cache.permute(2, 0, 1, 3))
else:
kv_cache = None
output = self.dense(context_layer)
return output, kv_cache
def chatglm2_attention_forward_8eb45c(
self, hidden_states, attention_mask, rotary_pos_emb, kv_cache=None, use_cache=True
):
# hidden_states: [sq, b, h]
# =================================================
# Pre-allocate memory for key-values for inference.
# =================================================
# =====================
# Query, Key, and Value
# =====================
# Attention heads [sq, b, h] --> [sq, b, (np * 3 * hn)]
device = hidden_states.device
mixed_x_layer = self.query_key_value(hidden_states)
if self.multi_query_attention:
(query_layer, key_layer, value_layer) = mixed_x_layer.split(
[
self.num_attention_heads_per_partition * self.hidden_size_per_attention_head,
self.num_multi_query_groups_per_partition * self.hidden_size_per_attention_head,
self.num_multi_query_groups_per_partition * self.hidden_size_per_attention_head,
],
dim=-1,
)
query_layer = query_layer.view(
query_layer.size()[:-1] + (self.num_attention_heads_per_partition,
self.hidden_size_per_attention_head)
)
key_layer = key_layer.view(
key_layer.size()[:-1] + (self.num_multi_query_groups_per_partition,
self.hidden_size_per_attention_head)
)
value_layer = value_layer.view(
value_layer.size()[:-1]
+ (self.num_multi_query_groups_per_partition, self.hidden_size_per_attention_head)
)
else:
new_tensor_shape = mixed_x_layer.size()[:-1] + (self.num_attention_heads_per_partition,
3 * self.hidden_size_per_attention_head)
mixed_x_layer = mixed_x_layer.view(*new_tensor_shape)
# [sq, b, np, 3 * hn] --> 3 [sq, b, np, hn]
(query_layer, key_layer, value_layer) = split_tensor_along_last_dim(mixed_x_layer, 3)
cur_length, batch_size = query_layer.shape[0], query_layer.shape[1]
# apply relative positional encoding (rotary embedding)
if rotary_pos_emb is not None:
if len(rotary_pos_emb) == 2 and isinstance(rotary_pos_emb, tuple):
# use_fuse_rope, see chatglm2_model_forward
cos, sin = rotary_pos_emb
rot_dim = cos.shape[-1]
query_layer = query_layer.transpose(0, 1)
key_layer = key_layer.transpose(0, 1)
query_layer_cur = query_layer[..., :rot_dim]
key_layer_cur = key_layer[..., :rot_dim]
# ipex_llm's apply_rotary_embedding can change the origin storage,
# so query_layer will get the result directly.
torch.ops.torch_ipex.apply_rotary_embedding(query_layer_cur, sin, cos, query_layer_cur)
torch.ops.torch_ipex.apply_rotary_embedding(key_layer_cur, sin, cos, key_layer_cur)
query_layer = query_layer.transpose(0, 1)
key_layer = key_layer.transpose(0, 1)
else:
query_layer = apply_rotary_pos_emb_chatglm(query_layer, rotary_pos_emb)
key_layer = apply_rotary_pos_emb_chatglm(key_layer, rotary_pos_emb)
if self.multi_query_attention:
if device.type == "xpu" and batch_size > 1: # use beam_search for generation.
# If batch_size > 1 on gpu, permute key/value_layer to [bs, np, sl, hn]
# to reduce memory usage. Otherwise，expend key/value_layer to [bs, nh, sl, hn].
key_layer = key_layer.permute(1, 2, 0, 3) # [bs, np, sl, hn]
value_layer = value_layer.permute(1, 2, 0, 3) # [bs, np, sl, hn]
else:
key_length = key_layer.size(0)
query_group_size = self.num_attention_heads_per_partition // \
self.num_multi_query_groups_per_partition
key_layer = key_layer.permute(1, 2, 0, 3).unsqueeze(-3) # [bs, nh/k, sl, hn]
key_layer = key_layer.expand(-1, -1, query_group_size, -1, -1)
key_layer = key_layer.contiguous().view((batch_size,
self.num_attention_heads_per_partition,
key_length,
self.hidden_size_per_attention_head))
value_layer = value_layer.permute(1, 2, 0, 3).unsqueeze(-3) # [bs, nh/k, sl, hn]
value_layer = value_layer.expand(-1, -1, query_group_size, -1, -1)
value_layer = value_layer.contiguous().view((batch_size,
self.num_attention_heads_per_partition,
key_length,
self.hidden_size_per_attention_head))
# adjust key and value for inference
if kv_cache is not None:
cache_k, cache_v = kv_cache
cache_k = cache_k.permute(1, 2, 0, 3)
cache_v = cache_v.permute(1, 2, 0, 3)
past_length = cache_k.size(2)
if cache_k.stride()[1] < (past_length + cur_length) * cache_k.size(3):
max_cache_length = past_length + cur_length + KV_CACHE_ALLOC_BLOCK_LENGTH
if device.type == "xpu" and batch_size > 1: # use beam_search for generation.
# If batch_size > 1 on gpu, use init_kv_cache to avoid empty cache for ensuring
# generation correctness.
# Set the num_heads in init_kv_cache to np, ensuring that the tensors of
# new_cache_k/v and key/value_layer have the same size.
new_cache_k, new_cache_v = init_kv_cache(batch_size,
self.num_multi_query_groups_per_partition,
self.hidden_size_per_attention_head,
past_length,
max_cache_length,
dtype=query_layer.dtype,
device=device)
else:
new_cache_k, new_cache_v = extend_kv_cache(batch_size,
self.num_attention_heads_per_partition,
self.hidden_size_per_attention_head,
past_length,
max_cache_length,
dtype=query_layer.dtype,
device=device)
new_cache_k[:] = cache_k
new_cache_v[:] = cache_v
cache_k = new_cache_k
cache_v = new_cache_v
key_layer, value_layer = append_kv_cache(cache_k, cache_v, key_layer, value_layer)
elif use_cache:
max_cache_length = max(KV_CACHE_ALLOC_MIN_LENGTH, cur_length) \
+ KV_CACHE_ALLOC_BLOCK_LENGTH
if device.type == "xpu" and batch_size > 1: # use beam_search for generation.
# Ensure the tensors of key/value_cache and key/value_layer have the same size.
nums_per_partition = self.num_multi_query_groups_per_partition
else:
nums_per_partition = self.num_attention_heads_per_partition
key_cache, value_cache = init_kv_cache(batch_size,
nums_per_partition,
self.hidden_size_per_attention_head,
cur_length,
max_cache_length,
dtype=query_layer.dtype,
device=device)
key_cache[:] = key_layer
value_cache[:] = value_layer
key_layer = key_cache
value_layer = value_cache
# If batch_size > 1, return tensors with shape [bs, np, sl, hn] as past_key_values. This could
# reduce memory usage as tensors are not expended to [bs, nh, sl, hn].
# Otherwise, return views of [bs, nh, sl, hn].
cache_key_layer = key_layer
cache_value_layer = value_layer
if use_cache:
kv_cache = (key_layer, value_layer)
else:
kv_cache = None
# ==================================
# core attention computation
# ==================================
if device.type == "xpu" and batch_size > 1: # use beam_search for generation.
# If batch_size > 1, expend key/value_layer to [ns, nh, sl, bn] for
# core attention computation.
# The expanded tensors will not be returned as past_key_values.
if self.multi_query_attention:
query_group_size = self.num_attention_heads_per_partition // \
self.num_multi_query_groups_per_partition
key_layer = key_layer.unsqueeze(-3)
key_layer = key_layer.expand(-1, -1, query_group_size, -1, -1)
save_length = key_layer.size(3)
# [bs, np, sl, hn] --> [bs, nh, sl, hn]
key_layer = key_layer.contiguous().view((batch_size,
self.num_attention_heads_per_partition,
save_length,
self.hidden_size_per_attention_head))
value_layer = value_layer.unsqueeze(-3)
value_layer = value_layer.expand(-1, -1, query_group_size, -1, -1)
# [bs, np, sl, hn] --> [bs, nh, sl, hn]
value_layer = value_layer.contiguous().view((batch_size,
self.num_attention_heads_per_partition,
save_length,
self.hidden_size_per_attention_head))
context_layer = core_attn_forward_8eb45c(query_layer, key_layer, value_layer, attention_mask)
# =================
# Output. [sq, b, h]
# =================
output = self.dense(context_layer)
return output, (cache_key_layer.permute(2, 0, 1, 3), cache_value_layer.permute(2, 0, 1, 3))
def core_attn_forward_8eb45c(query_layer, key_layer, value_layer, attention_mask):
query_layer = query_layer.permute(1, 2, 0, 3)
L, S = query_layer.shape[2], key_layer.shape[2]
batch_size, n_head, seq_len, head_dim = query_layer.shape
if attention_mask is None and L == S:
if should_split_qkv_tensor(query_layer, batch_size, n_head, seq_len):
# split second dim to block size = 8
block_size = 8
query_layer = query_layer.to(key_layer.dtype)
query_split = torch.split(query_layer, block_size, dim=1)
key_split = torch.split(key_layer, block_size, dim=1)
value_split = torch.split(value_layer, block_size, dim=1)
results = []
for q, k, v in zip(query_split, key_split, value_split):
result = glm_sdpa(q, k, v, is_causal=True)
results.append(result)
context_layer = torch.cat(results, dim=1)
else:
context_layer = glm_sdpa(query_layer,
key_layer,
value_layer,
is_causal=True)
else:
context_layer = glm_sdpa(query_layer,
key_layer,
value_layer,
attention_mask)
context_layer = context_layer.permute(2, 0, 1, 3)
new_context_layer_shape = context_layer.size()[:-2] + (-1,)
context_layer = context_layer.reshape(*new_context_layer_shape)
return context_layer
return output, past_key_value

2
python/llm/src/ipex_llm/transformers/models/utils.py
View file

@ -186,7 +186,7 @@ def apply_rotary_pos_emb(q, k, cos, sin, position_ids, model_family):
q_embed = (q * cos) + (rotate_half(q) * sin)
k_embed = (k * cos) + (rotate_half(k) * sin)
return q_embed, k_embed
elif model_family == "gptj":
elif model_family in ["gptj", "chatglm"]:
q_embed = (q * cos) + (rotate_every_two(q) * sin)
k_embed = (k * cos) + (rotate_every_two(k) * sin)
return q_embed, k_embed

8
python/llm/test/inference_gpu/test_transformers_api_attention.py
View file

@ -107,12 +107,6 @@ class Test_Optimize_Gpu_Model:
elif isinstance(t1, tuple) and isinstance(t2, tuple):
# if 'past_key_value'is of type tuple
for i, (t3, t4) in enumerate(zip(t1, t2)):
if model.config.architectures[0] == "ChatGLMModel" and \
hasattr(model.config, 'padded_vocab_size') and \
model.config.padded_vocab_size == 65024:
# chatglm2's past_key_value is expanded 16x for some speedup.
# We need to narrow it here.
t4 = t4[:, :, 15:17, :]
attn_output_diff.append(t3 - t4)
else:
# if 'past_key_value'is of type Cache, get last layer cache pair (key, value)
@ -171,7 +165,7 @@ class Test_Optimize_Gpu_Model:
# currently only need to compare the output of one self-attention layer.
layer_norm = "transformer.encoder.layers.27.input_layernorm"
self_attn = "transformer.encoder.layers.27.self_attention"
lower_bound = 8e-3
lower_bound = 4e-2
self.run_optimize_gpu_model(Name, Model, Tokenizer, model_path, self_attn, layer_norm, lower_bound)
def Mistral_gpu_model(self, Name, Model, Tokenizer, model_path):