Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Tests: gpt-oss dense-expert reload fixture drops replaced experts' quant state (#16246)

Daniel Han58cf8e9f58d7
diff --git a/tests/test_gpt_oss_bnb_checkpoint_layout.py b/tests/test_gpt_oss_bnb_checkpoint_layout.py
index 47caf9f55f7..9fd1ae1bb96 100644
--- a/tests/test_gpt_oss_bnb_checkpoint_layout.py
+++ b/tests/test_gpt_oss_bnb_checkpoint_layout.py
@@ -190,6 +190,44 @@ def _expert_weight(experts, name):
return w.detach().float().cpu()
+def _dense_expert_checkpoint(tensors_saved, original):
+ """Swap fused 4-bit experts for dense originals, dropping their quant state (else applied to dense bytes)."""
+ experts = ("experts.gate_up_proj", "experts.down_proj")
+ replaced = [name for name in tensors_saved if name.endswith(experts)]
+ tensors = {
+ name: value
+ for name, value in tensors_saved.items()
+ if not any(name.startswith(f"{r}.") for r in replaced)
+ }
+ for name in replaced:
+ tensors[name] = original[name]
+ return tensors
+
+
+def test_dense_expert_checkpoint_drops_replaced_quant_state():
+ w = torch.zeros(1)
+ q = f"{P}.experts.gate_up_proj"
+ saved = {
+ q: w,
+ f"{q}.absmax": w,
+ f"{q}.quant_map": w,
+ f"{q}.nested_absmax": w,
+ f"{q}.quant_state.bitsandbytes__nf4": w,
+ f"{P}.experts.gate_up_proj_bias": w,
+ "model.layers.0.self_attn.q_proj.weight": w,
+ "model.layers.0.self_attn.q_proj.weight.absmax": w,
+ }
+ dense = torch.ones(2, 2)
+ tensors = _dense_expert_checkpoint(saved, {q: dense})
+ assert tensors[q] is dense
+ assert set(tensors) == {
+ q,
+ f"{P}.experts.gate_up_proj_bias",
+ "model.layers.0.self_attn.q_proj.weight",
+ "model.layers.0.self_attn.q_proj.weight.absmax",
+ }
+
+
@pytest.mark.skipif(not has_real_cuda(), reason = "needs a CUDA GPU for bitsandbytes")
def test_reload_of_saved_fused_4bit_checkpoints(tmp_path):
pytest.importorskip("bitsandbytes")
@@ -216,6 +254,10 @@ def test_reload_of_saved_fused_4bit_checkpoints(tmp_path):
saved = str(tmp_path / "saved_4bit")
model.save_pretrained(saved)
tokenizer.save_pretrained(saved)
+ before = [
+ {name: _expert_weight(layer.mlp.experts, name) for name in ("gate_up_proj", "down_proj")}
+ for layer in model.model.layers
+ ]
del model
torch.cuda.empty_cache()
@@ -227,10 +269,7 @@ def test_reload_of_saved_fused_4bit_checkpoints(tmp_path):
for f in os.listdir(saved):
if f != "model.safetensors":
shutil.copy(os.path.join(saved, f), dense)
- tensors = dict(tensors_saved)
- for name in list(tensors):
- if name.endswith(("experts.gate_up_proj", "experts.down_proj")):
- tensors[name] = original[name]
+ tensors = _dense_expert_checkpoint(tensors_saved, original)
save_file(tensors, os.path.join(dense, "model.safetensors"), metadata = {"format": "pt"})
model, _ = FastLanguageModel.from_pretrained(dense, **load)
@@ -262,6 +301,10 @@ def test_reload_of_saved_fused_4bit_checkpoints(tmp_path):
else:
model, _ = FastLanguageModel.from_pretrained(saved, **load)
assert not hasattr(model.model.layers[0].mlp.experts, "gate_up_projs")
+ for layer_idx, layer in enumerate(model.model.layers):
+ for name, want in before[layer_idx].items():
+ got = _expert_weight(layer.mlp.experts, name)
+ assert torch.equal(got, want), (layer_idx, name)
del model