Skip to content

gguf: include MTP tensors for Qwen3-Next and Qwen3.5 models #20533

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

Closed
jesseposner wants to merge 1 commit into from
Closed

gguf: include MTP tensors for Qwen3-Next and Qwen3.5 models #20533

Show file tree
Hide file tree
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
44 changes: 42 additions & 2 deletions convert_hf_to_gguf.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4662,9 +4662,49 @@ def set_gguf_parameters(self):
rope_dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.hparams.get("partial_rotary_factor", 0.25)))

# MTP (multi-token prediction) support
n_mtp = self.hparams.get("mtp_num_hidden_layers", 0)
if n_mtp > 0:
self.block_count = self.hparams["num_hidden_layers"] + n_mtp
self.gguf_writer.add_block_count(self.block_count)
self.gguf_writer.add_nextn_predict_layers(n_mtp)

_mtp_remapper = {
"mtp.fc": "model.layers.{bid}.eh_proj",
"mtp.pre_fc_norm_embedding": "model.layers.{bid}.enorm",
"mtp.pre_fc_norm_hidden": "model.layers.{bid}.hnorm",
"mtp.norm": "model.layers.{bid}.shared_head.norm",
}

def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
if name.startswith("mtp"):
return # ignore MTP layers for now
if name.startswith("mtp."):
n_main = self.hparams["num_hidden_layers"]
n_mtp = self.hparams.get("mtp_num_hidden_layers", 0)
if n_mtp == 0:
return # MTP not configured

if "layers." in name:
# mtp.layers.{i}.* -> model.layers.{n_main + i}.*
import re
m = re.match(r"mtp\.layers\.(\d+)\.(.*)", name)
if m:
mtp_idx = int(m.group(1))
rest = m.group(2)
name = f"model.layers.{n_main + mtp_idx}.{rest}"
bid = n_main + mtp_idx
else:
# Non-layer MTP tensors (fc, norms): map to NEXTN names
stem = name.rsplit(".", 1)[0] # e.g. "mtp.fc"
suffix = "." + name.rsplit(".", 1)[1] # e.g. ".weight"
if stem in self._mtp_remapper:
template = self._mtp_remapper[stem]
for mtp_bid in range(n_main, n_main + n_mtp):
new_name = template.format(bid=mtp_bid) + suffix
yield from super().modify_tensors(data_torch, new_name, mtp_bid)
return
else:
logger.warning(f"Unknown MTP tensor: {name}")
return
if name.endswith(".A_log"):
data_torch = -torch.exp(data_torch)
elif name.endswith(".dt_bias"):
Expand Down
14 changes: 14 additions & 0 deletions src/llama-arch.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1024,6 +1024,13 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
LLM_TENSOR_SSM_IN,
LLM_TENSOR_SSM_NORM,
LLM_TENSOR_SSM_OUT,
LLM_TENSOR_FFN_NORM,
LLM_TENSOR_NEXTN_EH_PROJ,
LLM_TENSOR_NEXTN_EMBED_TOKENS,
LLM_TENSOR_NEXTN_ENORM,
LLM_TENSOR_NEXTN_HNORM,
LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,
LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,
};
case LLM_ARCH_QWEN35:
return {
Expand Down Expand Up @@ -1082,6 +1089,13 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
LLM_TENSOR_SSM_ALPHA,
LLM_TENSOR_SSM_NORM,
LLM_TENSOR_SSM_OUT,
LLM_TENSOR_FFN_NORM,
LLM_TENSOR_NEXTN_EH_PROJ,
LLM_TENSOR_NEXTN_EMBED_TOKENS,
LLM_TENSOR_NEXTN_ENORM,
LLM_TENSOR_NEXTN_HNORM,
LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,
LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,
};
case LLM_ARCH_QWEN3VL:
case LLM_ARCH_CHAMELEON:
Expand Down