-
Notifications
You must be signed in to change notification settings - Fork 10.1k
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.
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
add solar pro support #9541
add solar pro support #9541
Conversation
solar pro introduces block skip connections where blocks are connected to other, non-sequential blocks with a scale multiple this change adds 4 new keys to store the skip connections and one new tensor to store the scalar. the scalar is implemented a 1-dimensional tensor with 2 elements dervied from the model's bskcn_tv configuration. in general, the values are (bskcn_tv, 1 - bskcn_tv)
@@ -2538,6 +2565,14 @@ struct llama_hparams { | |||
return ssm_d_state * ssm_d_inner; | |||
} | |||
} | |||
|
|||
bool n_bskcn(uint32_t n, uint32_t il = 0) const { |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The n_
prefix implies that this returns an integer, however it returns a boolean.
is this PR active and maintained? |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I agree with the changes.
def prepare_tensors(self): | ||
if bskcn_tv := self.find_hparam(['bskcn_tv'], optional=True): | ||
# use bskcn_tv[1] for inference since bskcn_tv[0] is for training | ||
self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.BSKCN_TV), np.array([bskcn_tv[1], 1 - bskcn_tv[1]], dtype=np.float32)) | ||
|
||
super().prepare_tensors() |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think this should override generate_extra_tensors
instead of prepare_tensors
. Otherwise LoRA conversion will not work properly, at least since #9396.
if (hparams.n_bskcn(2, il)) { | ||
inpSA = ggml_add( | ||
ctx0, | ||
ggml_mul(ctx0, bskcn_1, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, 0)), |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
bskcn_1
is not necessarily initialized here, because a model file could be crafted to make hparams.n_bskcn(2, il)
return true
while making hparams.n_bskcn(1, il)
always return false
.
for i, bskcn in enumerate(self.hparams[k] for k in self.hparams.keys() if k.startswith("bskcn_") and k != 'bskcn_tv'): | ||
# store the skip connections as a layer index where a non-zero value indicates a skip connection | ||
# this approach simplifies lookup at inference time | ||
self.gguf_writer.add_block_skip_connection(i, [1 if n in bskcn else 0 for n in range(self.block_count)]) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This assumes bskcn_{n}
are in the correct order in config.json
. Why not instead iterate them by their names?
@mxyng Is this PR still on? |
solar pro introduces block skip connections where blocks are connected to other, non-sequential blocks with a scale multiple
this change adds 4 new keys to store the skip connections and one new tensor to store the scalar. the scalar is implemented as a 1-dimensional tensor with 2 elements derived from the model's bskcn_tv configuration. in general, the values are (bskcn_tv, 1 - bskcn_tv)