Skip to content

[qwen3.5] compat transformers 5.4.0 #18

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
Jintao-Huang merged 2 commits into from
Apr 8, 2026
+12 −6
Merged

[qwen3.5] compat transformers 5.4.0 #18

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
670422e
compat transformers 5.4.0
Jintao-Huang Apr 7, 2026
a331ddc
fix
Jintao-Huang Apr 8, 2026
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
16 changes: 11 additions & 5 deletions src/mcore_bridge/bridge/gpt_bridge.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -645,6 +645,7 @@ def _set_moe_state(
hf_prefix: str,
layer_idx: int,
to_mcore: bool,
is_mtp: bool = False,
):
if to_mcore:
hf_state_dict = self._remove_prefix(hf_state_dict, hf_prefix)
Expand Down Expand Up @@ -689,15 +690,18 @@ def _set_moe_state(
layer_idx,
to_mcore,
ep_rank=ep_rank,
is_mtp=is_mtp,
))
if to_mcore:
hf_state_dict = {}
else:
hf_state_dict = self._add_prefix(hf_state_dict, hf_prefix)
return hf_state_dict

def _get_hf_experts_attr(self):
def _get_hf_experts_attr(self, is_mtp: bool = False):
# return hf_grouped, is_gate_up
if self.model_type == 'qwen3_5_moe' and not is_mtp:
return True, True
if self.model_type in {'glm4v_moe', 'kimi_vl', 'qwen3_omni_moe'} or self.llm_model_type in {
'qwen2_moe', 'qwen3_moe', 'deepseek_v2', 'deepseek_v3', 'kimi_k2', 'dots1', 'ernie4_5_moe', 'glm4_moe',
'glm4_moe_lite', 'minimax_m2', 'olmoe', 'qwen3_next', 'qwen3_5_moe', 'glm_moe_dsa', 'deepseek_v32'
Expand All @@ -723,6 +727,7 @@ def _set_mlp_state(
layer_idx: int,
to_mcore: bool,
ep_rank: Optional[int] = None,
is_mtp: bool = False,
):
if to_mcore:
hf_state_dict = self._remove_prefix(hf_state_dict, hf_prefix)
Expand All @@ -738,7 +743,7 @@ def _set_mlp_state(
is_gate_up = any('gate_up_proj' in k for k in hf_state_dict.keys())
# transformers 5.0 compatibility
if not to_mcore and is_expert:
hf_grouped, is_gate_up = self._get_hf_experts_attr()
hf_grouped, is_gate_up = self._get_hf_experts_attr(is_mtp)
need_transpose = False
if hf_grouped:
need_transpose = self._get_need_transpose()
Expand Down Expand Up @@ -1401,7 +1406,7 @@ def _set_layer_attn(self, mg_layer, hf_state_dict, layer_idx: int, to_mcore: boo
'input_layernorm.weight', to_mcore)
return hf_state_dict

def _set_layer_mlp(self, mg_layer, hf_state_dict, layer_idx: int, to_mcore: bool):
def _set_layer_mlp(self, mg_layer, hf_state_dict, layer_idx: int, to_mcore: bool, is_mtp: bool = False):
mg_mlp = None if mg_layer is None else mg_layer.mlp
is_moe = True if hasattr(mg_mlp, 'experts') else False
if not to_mcore:
Expand All @@ -1410,7 +1415,8 @@ def _set_layer_mlp(self, mg_layer, hf_state_dict, layer_idx: int, to_mcore: bool
dist.all_reduce(is_moe, group=self.pp_group)
if is_moe:
hf_state_dict.update(
self._set_moe_state(mg_mlp, hf_state_dict, f'{self.hf_mlp_prefix}.', layer_idx, to_mcore))
self._set_moe_state(
mg_mlp, hf_state_dict, f'{self.hf_mlp_prefix}.', layer_idx, to_mcore, is_mtp=is_mtp))
self._set_state_dict(mg_layer, 'pre_mlp_layernorm.weight', hf_state_dict, 'post_attention_layernorm.weight',
to_mcore)
else:
Expand Down Expand Up @@ -1608,7 +1614,7 @@ def _convert_mtp_layer(self, lm_model, hf_state_dict, hf_prefix: str, layer_idx:
self._set_state_dict(lm_model, 'output_layer.weight', hf_state_dict, 'shared_head.head.weight',
to_mcore)
hf_state_dict.update(self._set_layer_attn(transformer_layer, hf_state_dict, -1, to_mcore))
hf_state_dict.update(self._set_layer_mlp(transformer_layer, hf_state_dict, -1, to_mcore))
hf_state_dict.update(self._set_layer_mlp(transformer_layer, hf_state_dict, -1, to_mcore, is_mtp=True))
if to_mcore:
hf_state_dict = {}
else:
Expand Down
2 changes: 1 addition & 1 deletion src/mcore_bridge/model/gpt_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -402,7 +402,7 @@ def _postprocess(
if self.config.is_multimodal and self.config.context_parallel_size > 1:
input_ids = split_cp_inputs(input_ids, getattr(packed_seq_params, 'cu_seqlens_q', None), 1)

if self.mtp_process:
if self.mtp_process and labels is not None:
if self.config.is_multimodal:
embedding_ = (self.embedding, decoder_input)
else:
Expand Down
Loading