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

Phi3 support compresskv (#11733)

Browse source 
* phi3 support compresskv

* fix phi3 mtl error

* fix conflict with quant kv

* fix abnormal on mtl

* fix style

* use slide windows size to compress kv

* support sliding window

* fix style

* fix style

* temp: partial support quant kv

* support quant kv with compress kv, todo: model check

* temp

* fix style

* fix style

* remove prepare

* address comment

* default -> 1.8k
This commit is contained in:
Yina Chen 2024-08-09 10:43:43 +03:00 committed by GitHub
parent d8808cc2e3
commit dd46c141bd
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 146 additions and 82 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

181
python/llm/src/ipex_llm/transformers/kv.py
View file

@ -155,62 +155,71 @@ def compress_kv(attn_config, key_states, query_states, value_states, attention_m
if q_len <= attn_config.max_capacity_prompt: if q_len <= attn_config.max_capacity_prompt:
return key_states, value_states return key_states, value_states
else: else:
key_states_expand = repeat_kv(key_states, num_key_value_groups).to(key_states.device) sliding_window_size = getattr(attn_config, "sliding_window", None)
attn_weights = torch.matmul(query_states[..., -attn_config.window_size:, :], if sliding_window_size is not None and sliding_window_size <= 2500:
key_states_expand.transpose(2, 3)) / math.sqrt(head_dim) return key_states[:, :, -sliding_window_size:, :], \
mask = torch.full((attn_config.window_size, attn_config.window_size), value_states[:, :, -sliding_window_size:, :]
torch.finfo(attn_weights.dtype).min,
device=attn_weights.device)
mask_cond = torch.arange(mask.size(-1), device=attn_weights.device)
mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
mask = mask.to(attn_weights.device)
attention_mask = mask[None, None, :, :]
attn_weights[:, :, -attn_config.window_size:, -attn_config.window_size:] += attention_mask
attn_weights = nn.functional.softmax(attn_weights, dim=-1,
dtype=torch.float32).to(query_states.dtype)
attn_weights_sum = attn_weights[:, :, -attn_config.window_size:,
:-attn_config.window_size].sum(dim=-2)
if attn_config.pooling == 'avgpool':
if num_key_value_groups > 1:
attn_cache = F.avg_pool2d(attn_weights_sum, kernel_size=(num_key_value_groups,
attn_config.kernel_size),
padding=(0, attn_config.kernel_size//2),
stride=(num_key_value_groups, 1))
else:
attn_cache = F.avg_pool1d(attn_weights_sum, kernel_size=attn_config.kernel_size,
padding=attn_config.kernel_size//2, stride=1)
elif attn_config.pooling == 'maxpool':
if num_key_value_groups > 1:
attn_cache = F.max_pool2d(attn_weights_sum,
kernel_size=(num_key_value_groups,
attn_config.kernel_size),
padding=(0, attn_config.kernel_size//2),
stride=(num_key_value_groups, 1))
else:
attn_cache = F.max_pool1d(attn_weights_sum, kernel_size=attn_config.kernel_size,
padding=attn_config.kernel_size//2, stride=1)
else: else:
invalidInputError(False, 'Pooling method not supported') key_states_expand = repeat_kv(key_states, num_key_value_groups).to(key_states.device)
indices = attn_cache.topk(attn_config.max_capacity_prompt - attn_config.window_size, attn_weights = torch.matmul(query_states[..., -attn_config.window_size:, :],
dim=-1).indices key_states_expand.transpose(2, 3)) / math.sqrt(head_dim)
indices = indices.unsqueeze(-1).expand(-1, -1, -1, head_dim) mask = torch.full((attn_config.window_size, attn_config.window_size),
k_past_compress = key_states[:, :, :-attn_config.window_size, :].gather(dim=2, torch.finfo(attn_weights.dtype).min,
index=indices) device=attn_weights.device)
v_past_compress = value_states[:, :, :-attn_config.window_size, :].gather(dim=2, mask_cond = torch.arange(mask.size(-1), device=attn_weights.device)
index=indices) mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
k_cur = key_states[:, :, -attn_config.window_size:, :] mask = mask.to(attn_weights.device)
v_cur = value_states[:, :, -attn_config.window_size:, :] attention_mask = mask[None, None, :, :]
key_states = torch.cat([k_past_compress, k_cur], dim=2)
value_states = torch.cat([v_past_compress, v_cur], dim=2) attn_weights[:, :, -attn_config.window_size:,
return key_states, value_states -attn_config.window_size:] += attention_mask
attn_weights = nn.functional.softmax(attn_weights, dim=-1,
dtype=torch.float32).to(query_states.dtype)
attn_weights_sum = attn_weights[:, :, -attn_config.window_size:,
:-attn_config.window_size].sum(dim=-2)
if attn_config.pooling == 'avgpool':
if num_key_value_groups > 1:
attn_cache = F.avg_pool2d(attn_weights_sum,
kernel_size=(num_key_value_groups,
attn_config.kernel_size),
padding=(0, attn_config.kernel_size//2),
stride=(num_key_value_groups, 1))
else:
attn_cache = F.avg_pool1d(attn_weights_sum, kernel_size=attn_config.kernel_size,
padding=attn_config.kernel_size//2, stride=1)
elif attn_config.pooling == 'maxpool':
if num_key_value_groups > 1:
attn_cache = F.max_pool2d(attn_weights_sum,
kernel_size=(num_key_value_groups,
attn_config.kernel_size),
padding=(0, attn_config.kernel_size//2),
stride=(num_key_value_groups, 1))
else:
attn_cache = F.max_pool1d(attn_weights_sum, kernel_size=attn_config.kernel_size,
padding=attn_config.kernel_size//2, stride=1)
else:
invalidInputError(False, 'Pooling method not supported')
indices = attn_cache.topk(attn_config.max_capacity_prompt - attn_config.window_size,
dim=-1).indices
indices = indices.unsqueeze(-1).expand(-1, -1, -1, head_dim)
k_past_compress = key_states[:, :, :-attn_config.window_size, :]\
.gather(dim=2, index=indices)
v_past_compress = value_states[:, :, :-attn_config.window_size, :]\
.gather(dim=2, index=indices)
k_cur = key_states[:, :, -attn_config.window_size:, :]
v_cur = value_states[:, :, -attn_config.window_size:, :]
key_states = torch.cat([k_past_compress, k_cur], dim=2)
value_states = torch.cat([v_past_compress, v_cur], dim=2)
return key_states, value_states
class DynamicCompressCache(DynamicCache): class DynamicCompressCache(DynamicCache):
def __init__(self, *args, **kwargs): def __init__(self, quant_kv=False, *args, **kwargs):
super().__init__(*args, **kwargs) super().__init__(*args, **kwargs)
self.real_kv_len = 0 self.real_kv_len = 0
self.quant_kv = quant_kv
self.append_kv_func = append_fp8_kv_cache if quant_kv else append_kv_cache
def update_seen_tokens(self, layer_idx, q_len): def update_seen_tokens(self, layer_idx, q_len):
if layer_idx == 0: if layer_idx == 0:
@ -260,49 +269,62 @@ class DynamicCompressCache(DynamicCache):
self.key_cache.append(key_states_compress) self.key_cache.append(key_states_compress)
self.value_cache.append(value_states_compress) self.value_cache.append(value_states_compress)
k_cache_compressed, v_cache_compressed = init_kv_cache( if not self.quant_kv:
bsz, num_heads, head_dim, k_cache_compressed, v_cache_compressed = init_kv_cache(
0, key_states_compress.size(2) + KV_CACHE_ALLOC_BLOCK_LENGTH, bsz, num_heads, head_dim,
key_states.dtype, key_states.device 0, key_states_compress.size(2) + KV_CACHE_ALLOC_BLOCK_LENGTH,
) key_states.dtype, key_states.device
k_cache_compressed, v_cache_compressed = append_kv_cache( )
else:
k_cache_compressed, v_cache_compressed = init_fp8_kv_cache(
bsz, num_heads, seq_len, head_dim,
device=key_states.device,
)
k_cache_compressed, v_cache_compressed = self.append_kv_func(
k_cache_compressed, v_cache_compressed, k_cache_compressed, v_cache_compressed,
key_states_compress, value_states_compress) key_states_compress, value_states_compress)
self.key_cache[layer_idx] = k_cache_compressed self.key_cache[layer_idx] = k_cache_compressed
self.value_cache[layer_idx] = v_cache_compressed self.value_cache[layer_idx] = v_cache_compressed
if key_states.stride(2) != head_dim: if key_states.stride(2) != head_dim:
k_cache, v_cache = init_kv_cache( if not self.quant_kv:
bsz, num_heads, head_dim, k_cache, v_cache = init_kv_cache(
0, key_states.size(2), bsz, num_heads, head_dim,
key_states.dtype, key_states.device 0, key_states.size(2),
) key_states.dtype, key_states.device
k_cache, v_cache = append_kv_cache(k_cache, v_cache, key_states, value_states) )
else:
k_cache, v_cache = init_fp8_kv_cache(
bsz, num_heads, 0, head_dim, key_states.device
)
k_cache, v_cache = self.append_kv_func(k_cache, v_cache,
key_states, value_states)
return k_cache, v_cache return k_cache, v_cache
else: else:
return key_states, value_states return key_states, value_states
else: else:
cache_k = self.key_cache[layer_idx] cache_k = self.key_cache[layer_idx]
cache_v = self.value_cache[layer_idx] cache_v = self.value_cache[layer_idx]
if not enough_kv_room: if not enough_kv_room and not self.quant_kv:
# allocate new # allocate new
new_c_k, new_c_v = extend_kv_cache(bsz, new_c_k, new_c_v = extend_kv_cache(
num_heads, # Support GQA bsz,
head_dim, num_heads, # Support GQA
cache_k.size(2), head_dim,
cache_k.size(2) + KV_CACHE_ALLOC_BLOCK_LENGTH, cache_k.size(2),
dtype=cache_k.dtype, cache_k.size(2) + KV_CACHE_ALLOC_BLOCK_LENGTH,
device=query_states.device) dtype=cache_k.dtype,
device=query_states.device)
new_c_k[:] = cache_k new_c_k[:] = cache_k
new_c_v[:] = cache_v new_c_v[:] = cache_v
cache_k = new_c_k cache_k = new_c_k
cache_v = new_c_v cache_v = new_c_v
key_states, value_states = append_kv_cache(cache_k, key_states, value_states = self.append_kv_func(cache_k,
cache_v, cache_v,
key_states, key_states,
value_states) value_states)
# update past_key_value # update past_key_value
self.key_cache[layer_idx] = key_states self.key_cache[layer_idx] = key_states
@ -316,3 +338,14 @@ class DynamicCompressCache(DynamicCache):
if len(self.key_cache) <= layer_idx: if len(self.key_cache) <= layer_idx:
return 0 return 0
return self.real_kv_len return self.real_kv_len
@classmethod
def from_legacy_cache(cls, past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
quantize_kv: Optional[bool] = False) -> "DynamicCache":
"""Converts a cache in the legacy cache format into an equivalent `DynamicCache`."""
cache = cls(quantize_kv)
if past_key_values is not None:
for layer_idx in range(len(past_key_values)):
key_states, value_states = past_key_values[layer_idx]
cache.update(key_states, value_states, layer_idx)
return cache

45
python/llm/src/ipex_llm/transformers/models/phi3.py
View file

@ -31,6 +31,7 @@
# See the License for the specific language governing permissions and # See the License for the specific language governing permissions and
# limitations under the License. # limitations under the License.
import os
import math import math
import torch import torch
import warnings import warnings
@ -40,11 +41,13 @@ from ipex_llm.transformers.models.utils import should_use_fuse_rope, rotate_half
from ipex_llm.transformers.models.utils import mlp_fusion_check, SILU from ipex_llm.transformers.models.utils import mlp_fusion_check, SILU
from ipex_llm.transformers.models.utils import use_sdp, use_sdp_causal from ipex_llm.transformers.models.utils import use_sdp, use_sdp_causal
from ipex_llm.transformers.models.utils import use_quantize_kv_cache, restore_fp8_kv_cache from ipex_llm.transformers.models.utils import use_quantize_kv_cache, restore_fp8_kv_cache
from ipex_llm.transformers.kv import DynamicNormalCache, DynamicFp8Cache from ipex_llm.transformers.models.utils import should_use_compresskv, is_enough_kv_cache_room_4_36
from ipex_llm.transformers.kv import DynamicNormalCache, DynamicFp8Cache, DynamicCompressCache
from typing import Optional, Tuple, List from typing import Optional, Tuple, List
from transformers.models.phi.modeling_phi import repeat_kv from transformers.models.phi.modeling_phi import repeat_kv
from transformers.cache_utils import Cache from transformers.cache_utils import Cache
KV_CACHE_ALLOC_BLOCK_LENGTH = int(os.environ.get("KV_CACHE_ALLOC_BLOCK_LENGTH", 256))
def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1): def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
@ -94,6 +97,9 @@ def attention_forward(
bsz, q_len, _ = hidden_states.size() bsz, q_len, _ = hidden_states.size()
# [CompressKV]
use_compresskv = isinstance(past_key_value, DynamicCompressCache)
qkv = self.qkv_proj(hidden_states) qkv = self.qkv_proj(hidden_states)
qkv = qkv.view(bsz, q_len, self.num_heads + 2 * self.num_key_value_heads, self.head_dim) qkv = qkv.view(bsz, q_len, self.num_heads + 2 * self.num_key_value_heads, self.head_dim)
qkv = qkv.transpose(1, 2) qkv = qkv.transpose(1, 2)
@ -127,12 +133,26 @@ def attention_forward(
cos, sin, position_ids) cos, sin, position_ids)
if past_key_value is not None: if past_key_value is not None:
key_states, value_states = past_key_value.update(key_states, value_states, # [CompressKV]
self.layer_idx, None) if use_compresskv:
enough_kv_room = is_enough_kv_cache_room_4_36(past_key_value, self.layer_idx)
key_states, value_states = past_key_value.update(
key_states, value_states, self.layer_idx,
query_states, attention_mask, self.num_key_value_groups,
self.config, enough_kv_room, KV_CACHE_ALLOC_BLOCK_LENGTH)
else:
key_states, value_states = past_key_value.update(key_states, value_states,
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):
# [CompressKV]
if use_compresskv:
# print(attention_mask.shape)
context_len = key_states.size(2)
attention_mask = attention_mask[:, :, :, -context_len:]
import xe_addons import xe_addons
if isinstance(past_key_value, DynamicFp8Cache): if isinstance(past_key_value,
DynamicFp8Cache) or (use_compresskv and past_key_value.quant_kv):
attn_output = xe_addons.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:
@ -148,7 +168,8 @@ def attention_forward(
# attn_output = xe_addons.sdp_causal(query_states, key_states, # attn_output = xe_addons.sdp_causal(query_states, key_states,
# value_states, attention_mask) # value_states, attention_mask)
else: else:
if isinstance(past_key_value, DynamicFp8Cache): if isinstance(past_key_value,
DynamicFp8Cache) or (use_compresskv and past_key_value.quant_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,
query_states.dtype) query_states.dtype)
# repeat k/v heads if n_kv_heads < n_heads # repeat k/v heads if n_kv_heads < n_heads
@ -235,10 +256,20 @@ def phi3_model_forward_wrapper(origin_model_forward):
use_cache = use_cache if use_cache is not None else self.config.use_cache use_cache = use_cache if use_cache is not None else self.config.use_cache
input = input_ids if input_ids is not None else inputs_embeds input = input_ids if input_ids is not None else inputs_embeds
use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.down_proj, input) use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.down_proj, input)
use_compress_kv = should_use_compresskv(input, input.shape[-1])
if use_cache: if use_cache:
if use_quantize_kv and not isinstance(past_key_values, DynamicFp8Cache): if use_compress_kv and not isinstance(past_key_values,
DynamicCompressCache):
past_key_values = DynamicCompressCache.\
from_legacy_cache(past_key_values,
quantize_kv=use_quantize_kv)
if use_quantize_kv and not isinstance(past_key_values,
(DynamicFp8Cache, DynamicCompressCache)):
past_key_values = DynamicFp8Cache.from_legacy_cache(past_key_values) past_key_values = DynamicFp8Cache.from_legacy_cache(past_key_values)
if not use_quantize_kv and not isinstance(past_key_values, DynamicNormalCache): if not use_quantize_kv and not use_compress_kv and not isinstance(past_key_values,
(DynamicNormalCache,
DynamicCompressCache
)):
past_key_values = DynamicNormalCache.from_legacy_cache(past_key_values) past_key_values = DynamicNormalCache.from_legacy_cache(past_key_values)
return origin_model_forward( return origin_model_forward(
self=self, self=self,

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

@ -490,7 +490,7 @@ def should_use_compresskv(x: torch.Tensor, prompt_len: int):
if use_compress_kv is None: if use_compress_kv is None:
return ( return (
get_xpu_device_type(x) == "mtl" get_xpu_device_type(x) == "mtl"
and prompt_len >= 2500 and prompt_len >= 1800
and prompt_len <= 4500 and prompt_len <= 4500
) )
else: else: