Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

R1 Hybrid: Refine DeepseekV3Attention with ipex-llm kernel #12843

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

R1 Hybrid: Refine DeepseekV3Attention with ipex-llm kernel #12843

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
114 changes: 75 additions & 39 deletions python/llm/example/GPU/CPU-GPU-Hybrid-DeepSeek-R1-Inference/generate.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,6 +25,8 @@
import ipex_llm

from ipex_llm.transformers import AutoModelForCausalLM
from ipex_llm.utils.common.log4Error import invalidInputError
from ipex_llm.transformers.models.common import scaled_dot_product_attention
from transformers import AutoTokenizer, GenerationConfig
from transformers.cache_utils import Cache, DynamicCache

Expand Down Expand Up @@ -147,54 +149,88 @@ def hybrid_DeepseekV3Attention_forward(
"with a layer index."
)
kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
# ipex-llm modify: test ipex-llm mla kernel
if False:
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
query_states = q
key_states = torch.cat(
[k_nope, k_pe.expand([-1, self.num_heads, -1, -1])],
dim=-1
)
import xe_addons
# print(self.rotary_emb.__class__.__name__)

if self.rotary_emb.__class__.__name__ == "DeepseekV3YarnRotaryEmbedding":
xe_addons.rotary_half_inplaced(self.rotary_emb.inv_freq, position_ids,
query_states[:, :, :, self.qk_nope_head_dim:],
key_states[:, :, :, self.qk_nope_head_dim:])
else:
invalidInputError(False, f"unknown rope method: {self.rotary_emb.__class__.__name__}")

if past_key_value is not None:
cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
key_states, value_states = past_key_value.update(
key_states, value_states, self.layer_idx, cache_kwargs
)

q_pe, k_pe = apply_rotary_pos_emb(q_pe, k_pe, cos, sin, position_ids)
else:
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)

query_states = k_pe.new_empty(bsz, self.num_heads, q_len, self.q_head_dim)
query_states[:, :, :, : self.qk_nope_head_dim] = q_nope
query_states[:, :, :, self.qk_nope_head_dim :] = q_pe
q_pe, k_pe = apply_rotary_pos_emb(q_pe, k_pe, cos, sin, position_ids)

key_states = k_pe.new_empty(bsz, self.num_heads, q_len, self.q_head_dim)
key_states[:, :, :, : self.qk_nope_head_dim] = k_nope
key_states[:, :, :, self.qk_nope_head_dim :] = k_pe
if past_key_value is not None:
cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
key_states, value_states = past_key_value.update(
key_states, value_states, self.layer_idx, cache_kwargs
)
query_states = k_pe.new_empty(bsz, self.num_heads, q_len, self.q_head_dim)
query_states[:, :, :, : self.qk_nope_head_dim] = q_nope
query_states[:, :, :, self.qk_nope_head_dim :] = q_pe

attn_weights = (
torch.matmul(query_states, key_states.transpose(2, 3)) * self.softmax_scale
)
key_states = k_pe.new_empty(bsz, self.num_heads, q_len, self.q_head_dim)
key_states[:, :, :, : self.qk_nope_head_dim] = k_nope
key_states[:, :, :, self.qk_nope_head_dim :] = k_pe

if past_key_value is not None:
cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
key_states, value_states = past_key_value.update(
key_states, value_states, self.layer_idx, cache_kwargs
)

if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
raise ValueError(
f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
f" {attn_weights.size()}"
if True:
attn_weights = None
attn_output = scaled_dot_product_attention(
query_states, key_states, value_states,
attention_mask, q_len == kv_seq_len, self.softmax_scale
)
assert attention_mask is not None
if attention_mask is not None:
if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
attn_output = attn_output[:, :, :, :self.v_head_dim]
else:
attn_weights = (
torch.matmul(query_states, key_states.transpose(2, 3)) * self.softmax_scale
)

if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
raise ValueError(
f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
f" {attn_weights.size()}"
)
attn_weights = attn_weights + attention_mask

# upcast attention to fp32
attn_weights = nn.functional.softmax(
attn_weights, dim=-1, dtype=torch.float32
).to(query_states.dtype)
attn_weights = nn.functional.dropout(
attn_weights, p=self.attention_dropout, training=self.training
)
attn_output = torch.matmul(attn_weights, value_states)

if attn_output.size() != (bsz, self.num_heads, q_len, self.v_head_dim):
raise ValueError(
f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.v_head_dim)}, but is"
f" {attn_output.size()}"
assert attention_mask is not None
if attention_mask is not None:
if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
raise ValueError(
f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
)
attn_weights = attn_weights + attention_mask

# upcast attention to fp32
attn_weights = nn.functional.softmax(
attn_weights, dim=-1, dtype=torch.float32
).to(query_states.dtype)
attn_weights = nn.functional.dropout(
attn_weights, p=self.attention_dropout, training=self.training
)
attn_output = torch.matmul(attn_weights, value_states)

if attn_output.size() != (bsz, self.num_heads, q_len, self.v_head_dim):
raise ValueError(
f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.v_head_dim)}, but is"
f" {attn_output.size()}"
)

attn_output = attn_output.transpose(1, 2).contiguous()

Expand Down