Skip to content

convert : set expert gating func in base class #17279

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.

Sign up for GitHub

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

Merged
merged 1 commit into from
Nov 15, 2025
Merged

convert : set expert gating func in base class #17279

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
43 changes: 9 additions & 34 deletions convert_hf_to_gguf.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -825,6 +825,15 @@ def set_gguf_parameters(self):
self.gguf_writer.add_expert_group_used_count(n_group_used)
logger.info(f"gguf: expert groups used count = {n_group_used}")

if (score_func := self.find_hparam(["score_function", "scoring_func", "score_func"], optional=True)) is not None:
Copy link
Collaborator

@ngxson ngxson Nov 15, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm wondering if there is a way to mark this as required (optional=False) for certain model

Copy link
Collaborator Author

@CISC CISC Nov 15, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Then you'd have to maintain a list, not very practical, besides if your model requires this it will throw at graph build.

if score_func == "sigmoid":
self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
elif score_func == "softmax":
self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SOFTMAX)
else:
raise ValueError(f"Unsupported expert score gating function value: {score_func}")
logger.info(f"gguf: expert score gating function = {score_func}")

if (head_dim := self.hparams.get("head_dim")) is not None:
self.gguf_writer.add_key_length(head_dim)
self.gguf_writer.add_value_length(head_dim)
Expand Down Expand Up @@ -2553,15 +2562,6 @@ def set_gguf_parameters(self):
if (n_dense_layers := self.hparams.get("num_dense_layers")) is not None:
self.gguf_writer.add_leading_dense_block_count(n_dense_layers)

# Expert Gating Function
score_func = self.hparams.get("score_func")
if score_func == "sigmoid":
self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
elif score_func == "softmax":
self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SOFTMAX)
elif score_func is not None:
raise ValueError(f"Unsupported score_function value: {score_func}")

# Route normalization and scaling
if (route_norm := self.hparams.get("route_norm")) is not None:
self.gguf_writer.add_expert_weights_norm(route_norm)
Expand Down Expand Up @@ -7182,13 +7182,6 @@ def set_gguf_parameters(self):
self.gguf_writer.add_expert_weights_scale(hparams["routed_scaling_factor"])
self.gguf_writer.add_expert_weights_norm(hparams["norm_topk_prob"])

if hparams["scoring_func"] == "sigmoid":
self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
elif hparams["scoring_func"] == "softmax":
self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SOFTMAX)
else:
raise ValueError(f"Unsupported scoring_func value: {hparams['scoring_func']}")

self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])

rope_scaling = self.hparams.get("rope_scaling") or {}
Expand Down Expand Up @@ -7294,12 +7287,6 @@ def __init__(self, *args, **kwargs):

def set_gguf_parameters(self):
super().set_gguf_parameters()
if self.hparams["scoring_func"] == "sigmoid":
self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
elif self.hparams["scoring_func"] == "softmax":
self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SOFTMAX)
else:
raise ValueError(f"Unsupported scoring_func value: {self.hparams['scoring_func']}")

self.gguf_writer.add_expert_feed_forward_length(self.find_hparam(["intermediate_size"]))
self.gguf_writer.add_rope_dimension_count(self.find_hparam(["rotary_dim"]))
Expand Down Expand Up @@ -7392,11 +7379,6 @@ def set_gguf_parameters(self):
self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])
self.gguf_writer.add_expert_weights_norm(self.hparams["norm_topk_prob"])

if self.hparams["scoring_func"] == "noaux_tc":
Copy link
Collaborator

@ngxson ngxson Nov 15, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This noaux_tc is probably not handled now

Copy link
Collaborator Author

@CISC CISC Nov 15, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, that's on purpose, as mentioned in OP, this was incorrectly set in dots1, and fixed shortly after release, but never here.

self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
else:
raise ValueError(f"Unsupported scoring_func value: {self.hparams['scoring_func']}")

def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
if name.endswith("e_score_correction_bias"):
name = name.replace("e_score_correction_bias", "e_score_correction.bias")
Expand Down Expand Up @@ -8717,13 +8699,6 @@ def set_gguf_parameters(self):
self.gguf_writer.add_expert_shared_count(hparams["num_shared_experts"])
self.gguf_writer.add_expert_weights_norm(hparams["norm_topk_prob"])

if hparams["score_function"] == "sigmoid":
self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
elif hparams["score_function"] == "softmax":
self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SOFTMAX)
else:
raise ValueError(f"Unsupported score_function value: {hparams['score_function']}")

if (nextn_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
self.gguf_writer.add_nextn_predict_layers(nextn_layers)

Expand Down