huggingface/text-generation-inference
5
0
Fork 0
mirror of https://github.com/huggingface/text-generation-inference.git synced 2025-09-10 11:54:52 +00:00
Code Issues Packages Projects Releases Wiki Activity

working

Browse Source
This commit is contained in:
OlivierDehaene 2023-05-30 14:45:31 +02:00
parent 73cf93f1ee
commit bbb1d9e704
1 changed files with 55 additions and 29 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

84
server/text_generation_server/models/custom_modeling/flash_rw_modeling.py
View File

@ -257,6 +257,10 @@ class FlashRWLargeAttention(torch.nn.Module):
self.num_groups = self.num_groups // process_group.size() self.num_groups = self.num_groups // process_group.size()
self.num_heads_config = num_heads
self.num_heads_kv_config = num_heads_kv
self.num_groups = 64
def forward( def forward(
self, self,
hidden_states, hidden_states,
@ -268,37 +272,56 @@ class FlashRWLargeAttention(torch.nn.Module):
layer_past_present_indices, layer_past_present_indices,
cu_seqlens_q, cu_seqlens_q,
): ):
cu_shape = hidden_states.shape[0]
qkv = self.query_key_value(hidden_states) qkv = self.query_key_value(hidden_states)
qkv = qkv.view(-1, self.num_groups, self.num_heads + 2, self.head_size) qkv = qkv.view(cu_shape, -1, self.num_heads_config // self.num_heads_kv_config +2, 64)
q = qkv[:, :, :-2]
k = qkv[:, :, [-2]]
v = qkv[:, :, [-1]]
# Split query from key_value k = torch.broadcast_to(k, q.shape)
query, kv = qkv.split( v = torch.broadcast_to(v, q.shape)
[self.num_heads, 2],
dim=2,
)
# Prepare query and key_value for indexing q = q.reshape(cu_shape, -1, self.head_size)
query = query.reshape(-1, self.num_groups * self.num_heads, self.head_size) k = k.reshape(cu_shape, -1, self.head_size)
kv = kv.transpose(1, 2) v = v.reshape(cu_shape, -1, self.head_size)
logger.error(k.shape)
# qkv = qkv.view(-1, self.num_groups, self.num_heads + 2, self.head_size)
#
# # Split query from key_value
# query, kv = qkv.split(
# [self.num_heads, 2],
# dim=2,
# )
#
# # Prepare query and key_value for indexing
# query = query.reshape(-1, self.num_groups * self.num_heads, self.head_size)
# kv = kv.transpose(1, 2)
# Inplace rotary # Inplace rotary
self.rotary_emb(query, cos, sin) self.rotary_emb(q, cos, sin)
self.rotary_emb(kv[:, 0], cos, sin) self.rotary_emb(k, cos, sin)
# Prefill # Prefill
if layer_past_present_indices is None: if layer_past_present_indices is None:
# Copy to layer past # Copy to layer past
layer_past[...] = kv # layer_past[...] = kv
k, v = kv.split(1, dim=1) # k, v = kv.split(1, dim=1)
# Expand to query shape # Expand to query shape
k = k.expand(-1, self.num_heads, self.num_groups, self.head_size).reshape(-1, self.num_groups * self.num_heads, self.head_size) # k = k.transpose(1, 2).expand(-1, self.num_groups, self.num_heads, self.head_size).reshape(-1, self.num_groups * self.num_heads, self.head_size)
v = v.expand(-1, self.num_heads, self.num_groups, self.head_size).reshape(-1, self.num_groups * self.num_heads, self.head_size) # v = v.transpose(1, 2).expand(-1, self.num_groups, self.num_heads, self.head_size).reshape(-1, self.num_groups * self.num_heads, self.head_size)
layer_past[:, 0] = k
layer_past[:, 1] = v
# output # output
attn_output = torch.empty_like(query) attn_output = torch.empty_like(q)
# flash attention # flash attention
flash_attn_cuda.fwd( flash_attn_cuda.fwd(
query, q,
k, k,
v, v,
attn_output, attn_output,
@ -317,19 +340,22 @@ class FlashRWLargeAttention(torch.nn.Module):
# Decode # Decode
else: else:
# Add present to the layer_past tensor at the correct indices # Add present to the layer_past tensor at the correct indices
layer_past[layer_past_present_indices] = kv # layer_past[layer_past_present_indices] = kv
k, v = layer_past.split(1, dim=1) # k, v = layer_past.split(1, dim=1)
# Expand to query shape # Expand to query shape
k = k.expand(-1, self.num_heads, self.num_groups, self.head_size).reshape(-1, self.num_groups * self.num_heads, self.head_size) # k = k.expand(-1, self.num_heads, self.num_groups, self.head_size).reshape(-1, self.num_groups * self.num_heads, self.head_size)
v = v.expand(-1, self.num_heads, self.num_groups, self.head_size).reshape(-1, self.num_groups * self.num_heads, self.head_size) # v = v.expand(-1, self.num_heads, self.num_groups, self.head_size).reshape(-1, self.num_groups * self.num_heads, self.head_size)
layer_past[layer_past_present_indices, 0] = k
layer_past[layer_past_present_indices, 1] = v
# output # output
attn_output = torch.empty_like(query) attn_output = torch.empty_like(q)
# flash attention # flash attention
flash_attn_cuda.fwd( flash_attn_cuda.fwd(
query, q,
k, layer_past[:, 0],
v, layer_past[:, 1],
attn_output, attn_output,
cu_seqlens_q, cu_seqlens_q,
cu_seqlens, cu_seqlens,
@ -344,7 +370,7 @@ class FlashRWLargeAttention(torch.nn.Module):
None, None,
) )
return self.dense(attn_output.view(-1, self.num_heads * self.num_groups * self.head_size)) return self.dense(attn_output.view(cu_shape, -1))
class FlashMLP(nn.Module): class FlashMLP(nn.Module):
@ -498,8 +524,8 @@ class FlashRWLargeLayer(nn.Module):
layer_past_present_indices, layer_past_present_indices,
cu_seqlens_q, cu_seqlens_q,
): ):
ln_attn, residual = self.ln_attn(hidden_states, residual) ln_attn, _ = self.ln_attn(hidden_states)
ln_mlp, _ = self.ln_mlp(hidden_states, residual) ln_mlp, _ = self.ln_mlp(hidden_states)
# Self attention. # Self attention.
attn_output = self.self_attention( attn_output = self.self_attention(
@ -522,7 +548,7 @@ class FlashRWLargeLayer(nn.Module):
if self.process_group is not None: if self.process_group is not None:
torch.distributed.all_reduce(intermediate, group=self.process_group) torch.distributed.all_reduce(intermediate, group=self.process_group)
return intermediate, residual return intermediate + hidden_states, None
class FlashRWPreTrainedModel(PreTrainedModel): class FlashRWPreTrainedModel(PreTrainedModel):