Skip to content

[model] fix: Preserve GLM-5 MTP layers during conversion #4960

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
yaoyu-33 wants to merge 1 commit into from
Closed

[model] fix: Preserve GLM-5 MTP layers during conversion #4960

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
4 changes: 0 additions & 4 deletions src/megatron/bridge/models/glm_moe_dsa/glm5_bridge.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -74,10 +74,6 @@ def provider_bridge(self, hf_pretrained: PreTrainedCausalLM) -> MLAModelProvider
provider.qk_layernorm = True
provider.multi_latent_attention = True

# Disable MTP (Multi-Token Prediction) by default
# HF config has num_nextn_predict_layers=1
provider.mtp_num_layers = None

provider.moe_grouped_gemm = True
provider.moe_router_pre_softmax = True
provider.moe_token_dispatcher_type = "alltoall"
Expand Down
10 changes: 9 additions & 1 deletion tests/unit_tests/models/glm_moe_dsa/test_glm5_bridge.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,6 +42,7 @@ def _provider_from_hf_config(monkeypatch: pytest.MonkeyPatch, **config_overrides
config = {
"first_k_dense_replace": 3,
"num_hidden_layers": 78,
"num_nextn_predict_layers": 1,
"moe_intermediate_size": 2048,
"n_shared_experts": 1,
"rope_parameters": {"rope_theta": 1_000_000},
Expand All @@ -55,11 +56,18 @@ def _provider_from_hf_config(monkeypatch: pytest.MonkeyPatch, **config_overrides
monkeypatch.setattr(
MegatronModelBridge,
"provider_bridge",
lambda _self, _hf_pretrained: SimpleNamespace(),
lambda _self, hf_pretrained: SimpleNamespace(mtp_num_layers=hf_pretrained.config.num_nextn_predict_layers),
)
return GLM5Bridge().provider_bridge(SimpleNamespace(config=SimpleNamespace(**config)))


def test_provider_bridge_preserves_mtp_architecture_from_hf_config(monkeypatch: pytest.MonkeyPatch) -> None:
"""GLM-5 conversion preserves the MTP layers declared by the checkpoint config."""
provider = _provider_from_hf_config(monkeypatch)

assert provider.mtp_num_layers == 1


@pytest.mark.parametrize(
("config_overrides", "expected_topk_freq", "expected_skip_topk_offset"),
[
Expand Down
Loading