ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Update llama.py

Browse source
This commit is contained in:
Daniel Han-Chen 2024-02-08 02:57:09 +11:00
commit 998097394a
1 changed files with 26 additions and 52 deletions
Download patch file Download diff file Expand all files Collapse all files

76
unsloth/models/llama.py
View file

@ -72,17 +72,8 @@ pass
from math import sqrt as math_sqrt
KV_CACHE_INCREMENT = 1024 # KV Cache update size
KV_CACHE_INCREMENT = 128 # KV Cache update size
@torch.compile(options = {
"epilogue_fusion" : True,
"max_autotune" : True,
"fallback_random" : False,
"shape_padding" : True,
"triton.cudagraphs" : False,
"trace.enabled" : True,
"trace.graph_diagram" : True,
}, dynamic = True,)
def LlamaAttention_fast_forward_inference(
self,
hidden_states: torch.Tensor,
@ -144,16 +135,16 @@ def LlamaAttention_fast_forward_inference(
self.RH_Q = torch.empty((bsz, n_heads, 1, head_dim), dtype = dtype, device = "cuda")
self.attention = torch.empty((bsz, n_heads, 1, KV_CACHE_INCREMENT+seq_len), dtype = dtype, device = "cuda")
self.scalar = 1.0 / math_sqrt(self.head_dim)
# elif kv_seq_len >= self.paged_attention.shape[0]:
# self.paged_attention.resize_((self.paged_attention.shape[0]+KV_CACHE_INCREMENT, 2, bsz, n_kv_heads, head_dim))
# self.paged_attention_K = self.paged_attention[:,0]
# self.paged_attention_V = self.paged_attention[:,1]
# self.attention.resize_((bsz, n_heads, 1, self.attention.shape[-1]+KV_CACHE_INCREMENT))
# pass
elif kv_seq_len >= self.paged_attention.shape[0]:
self.paged_attention.resize_((self.paged_attention.shape[0]+KV_CACHE_INCREMENT, 2, bsz, n_kv_heads, head_dim))
self.paged_attention_K = self.paged_attention[:,0]
self.paged_attention_V = self.paged_attention[:,1]
self.attention.resize_((bsz, n_heads, 1, self.attention.shape[-1]+KV_CACHE_INCREMENT))
pass
Qn = fast_linear_forward(self.q_proj, Xn)#, out = self.temp_QA[0])
Kn = fast_linear_forward(self.k_proj, Xn)#, out = self.temp_KV[0])
Vn = fast_linear_forward(self.v_proj, Xn)#, out = self.temp_KV[1])
Qn = fast_linear_forward(self.q_proj, Xn, out = self.temp_QA[0])
Kn = fast_linear_forward(self.k_proj, Xn, out = self.temp_KV[0])
Vn = fast_linear_forward(self.v_proj, Xn, out = self.temp_KV[1])
Qn = Qn.view(bsz, 1, n_heads, head_dim).transpose(1, 2)
Kn = Kn.view(bsz, 1, n_kv_heads, head_dim).transpose(1, 2)
Vn = Vn.view(bsz, 1, n_kv_heads, head_dim).transpose(1, 2)
@ -181,15 +172,14 @@ def LlamaAttention_fast_forward_inference(
Vn = self.paged_attention_V[:kv_seq_len].permute(1, 2, 0, 3)
# Handle sliding windows
attention_size = getattr(self.config, "sliding_window", kv_seq_len)
if kv_seq_len <= attention_size:
attention_size = kv_seq_len
Knn, Vnn = Kn, Vn
else:
sliding_window = getattr(self.config, "sliding_window", None)
if sliding_window is not None and kv_seq_len > sliding_window:
# From https://github.com/huggingface/transformers/blob/main/src/transformers/models/mistral/modeling_mistral.py#L193
slicing_tokens = 1 - attention_size
slicing_tokens = 1 - sliding_window
Knn = Kn[:, :, slicing_tokens:, :]#.contiguous()
Vnn = Vn[:, :, slicing_tokens:, :]#.contiguous()
else:
Knn, Vnn = Kn, Vn
pass
# Grouped query attention
@ -205,51 +195,35 @@ def LlamaAttention_fast_forward_inference(
# pass
# Attention
A = torch.matmul(Qn, Knn.transpose(2, 3)) #out = self.attention[:,:,:,:attention_size])
A *= 1.0 / (self.head_dim**0.5)
A = torch.matmul(Qn, Knn.transpose(2, 3), out = self.attention[:,:,:,:kv_seq_len])
A *= self.scalar
A[:] = torch.nn.functional.softmax(A, dim = -1, dtype = torch.float32)#.to(A.dtype)
A = torch.matmul(A, Vnn)#, out = Qn)
A = torch.matmul(A, Vnn, out = Qn)
A = A.transpose(1, 2)
A = A.reshape(bsz, 1, self.hidden_size)
A = fast_linear_forward(self.o_proj, A)#, out = self.temp_QA[1])
A = fast_linear_forward(self.o_proj, A, out = self.temp_QA[1])
return A, (Kn, Vn)
pass
@torch.compile(options = {
"epilogue_fusion" : True,
"max_autotune" : True,
"fallback_random" : False,
"shape_padding" : True,
"triton.cudagraphs" : False,
"trace.enabled" : True,
"trace.graph_diagram" : True,
}, dynamic = True,)
def fast_mlp_inference(self, X):
# gate = self.gate_proj(X)
# up = self.up_proj(X)
bsz, _, hd = X.shape
mlp_size = self.config.intermediate_size
#temp = torch.empty((2, bsz, 1, mlp_size), dtype = X.dtype, device = "cuda")
temp = torch.empty((2, bsz, 1, mlp_size), dtype = X.dtype, device = "cuda")
gate = fast_linear_forward(self.gate_proj, X)#, out = temp[0])
up = fast_linear_forward(self. up_proj, X)#, out = temp[1])
gate = fast_linear_forward(self.gate_proj, X, out = temp[0])
up = fast_linear_forward(self. up_proj, X, out = temp[1])
gate = torch.nn.functional.silu(gate, inplace = True)
gate *= up
# X = self.down_proj(gate)
down = fast_linear_forward(self.down_proj, gate)#, out = up[:,:,:hd])
down = fast_linear_forward(self.down_proj, gate, out = up[:,:,:hd])
return down
pass
@torch.compile(options = {
"epilogue_fusion" : True,
"max_autotune" : True,
"fallback_random" : False,
"shape_padding" : True,
"triton.cudagraphs" : False,
"trace.enabled" : True,
"trace.graph_diagram" : True,
}, dynamic = True,)
def fast_rms_layernorm_inference(self, X):
old_dtype = X.dtype
XX = X.to(torch.float32)