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

inference

Browse source
This commit is contained in:
Daniel Han-Chen 2024-02-01 19:46:38 +11:00
commit 8920cafbe7
2 changed files with 5 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

6
unsloth/kernels/utils.py
View file

@ -119,7 +119,7 @@ def fast_gemv(X, W, quant_state, out = None):
# For fast X @ W where seq_len == 1
# From https://github.com/TimDettmers/bitsandbytes/blob/main/bitsandbytes/functional.py#L1469
bsz, q_len, hd = X.shape
assert(q_len == 1)
# assert(q_len == 1)
if type(quant_state) is not list:
# https://github.com/TimDettmers/bitsandbytes/pull/763/files
@ -138,7 +138,7 @@ def fast_gemv(X, W, quant_state, out = None):
offset, state2 = compressed_stats
absmax2, code2, blocksize2, _, _, _, _ = state2
pass
assert(dtype == X.dtype)
# assert(dtype == X.dtype)
bout = shape[0]
if out is None:
@ -152,7 +152,7 @@ def fast_gemv(X, W, quant_state, out = None):
k = shape[1]
lda = shape[0]
ldc = shape[0]
ldb = (X.shape[-1]+1)//2
ldb = (hd+1)//2
m = ctypes.c_int32(m)
n = ctypes.c_int32(n)
k = ctypes.c_int32(k)

4
unsloth/models/llama.py
View file

@ -286,7 +286,7 @@ def fast_mlp_inference(self, X):
# up = self.up_proj(X)
bsz, _, hd = X.shape
mlp_size = self.config.intermediate_size
temp = torch.empty((3, 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])
@ -294,7 +294,7 @@ def fast_mlp_inference(self, X):
gate *= up
# X = self.down_proj(gate)
down = fast_linear_forward(self.down_proj, gate, out = temp[2])
down = fast_linear_forward(self.down_proj, gate)
return down
pass