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 02:30:49 +11:00
commit e094914b0e
1 changed files with 13 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

18
unsloth/models/llama.py
View file

@ -206,21 +206,29 @@ def LlamaAttention_fast_forward_inference(
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