Fix HF PTQ empty-init dtype kwargs

examples/hf_ptq/example_utils.py

@@ -731,14 +731,25 @@ def has_pack_quantized_config(config):
             auto_model_module = getattr(transformers, architecture)
             from_config = auto_model_module._from_config
 
+        is_decilm = "DeciLM" in architecture
         with init_empty_weights(include_buffers=True):
             # When computing the device_map, assuming bfloat16 precision by default,
             # unless specified by the hf_config.
-            torch_dtype = getattr(hf_config, "torch_dtype", torch.bfloat16)
+            config_dtype = (
+                getattr(hf_config, "dtype", None)
+                or getattr(hf_config, "torch_dtype", None)
+                or torch.bfloat16
+            )
+            if isinstance(config_dtype, str):
+                config_dtype = getattr(torch, config_dtype)
             model_kwargs2 = model_kwargs.copy()
             if auto_model_module not in [AutoModelForCausalLM, AutoModel]:
                 model_kwargs2.pop("trust_remote_code", None)
-            model_kwargs2["dtype"] = torch_dtype
+            if is_decilm:
+                model_kwargs2["torch_dtype"] = config_dtype
+                model_kwargs2.pop("dtype", None)
+            else:
+                model_kwargs2["dtype"] = config_dtype
             model_kwargs2.pop("max_memory", None)
             model = from_config(hf_config, **model_kwargs2)
 
@@ -760,10 +771,13 @@ def has_pack_quantized_config(config):
             )
             model_kwargs["max_memory"] = max_memory
 
+        model_kwargs2 = model_kwargs.copy()
+        if is_decilm:
+            model_kwargs2.pop("dtype", None)
         model = auto_model_module.from_pretrained(
             ckpt_path,
             device_map=device_map,
-            **model_kwargs,
+            **model_kwargs2,
         )
     model.eval()
     if has_pack_quantized_config(hf_config):

tests/examples/hf_ptq/test_example_utils.py

@@ -19,8 +19,10 @@
 """
 
 import json
+from contextlib import nullcontext
 from types import SimpleNamespace
 
+import pytest
 import torch
 from _test_utils.examples.hf_ptq_example_utils import example_utils
 from safetensors.torch import save_file
@@ -194,3 +196,94 @@ def test_get_original_hf_quant_method_none_for_unquantized():
         example_utils.get_original_hf_quant_method(SimpleNamespace(quantization_config=None))
         is None
     )
+
+
+@pytest.mark.parametrize(
+    (
+        "architecture",
+        "model_class_name",
+        "expected_config_dtype_kwarg",
+        "unexpected_config_dtype_kwarg",
+    ),
+    [
+        ("DeciLMForCausalLM", "AutoModelForCausalLM", "torch_dtype", "dtype"),
+        ("LlamaForCausalLM", "LlamaForCausalLM", "dtype", "torch_dtype"),
+    ],
+)
+def test_get_model_uses_expected_dtype_kwarg(
+    monkeypatch,
+    architecture,
+    model_class_name,
+    expected_config_dtype_kwarg,
+    unexpected_config_dtype_kwarg,
+):
+    calls = {}
+    hf_config = SimpleNamespace(
+        architectures=[architecture],
+        dtype=torch.float16,
+        model_type="llama",
+        torch_dtype=torch.bfloat16,
+    )
+
+    class FakeModel:
+        def eval(self):
+            calls["eval"] = True
+
+    class FakeAutoModelForCausalLM:
+        @staticmethod
+        def from_config(config, **kwargs):
+            calls["from_config"] = kwargs
+            assert config is hf_config
+            assert kwargs[expected_config_dtype_kwarg] is torch.float16
+            assert unexpected_config_dtype_kwarg not in kwargs
+            assert "max_memory" not in kwargs
+            return FakeModel()
+
+        @staticmethod
+        def from_pretrained(*args, **kwargs):
+            calls["from_pretrained"] = kwargs
+            assert "dtype" not in kwargs
+            assert "torch_dtype" not in kwargs
+            return FakeModel()
+
+    class FakeLlamaForCausalLM:
+        @staticmethod
+        def _from_config(config, **kwargs):
+            calls["from_config"] = kwargs
+            assert config is hf_config
+            assert kwargs[expected_config_dtype_kwarg] is torch.float16
+            assert unexpected_config_dtype_kwarg not in kwargs
+            assert "max_memory" not in kwargs
+            return FakeModel()
+
+        @staticmethod
+        def from_pretrained(*args, **kwargs):
+            calls["from_pretrained"] = kwargs
+            assert kwargs["dtype"] == "auto"
+            assert "torch_dtype" not in kwargs
+            return FakeModel()
+
+    monkeypatch.setattr(
+        example_utils.AutoConfig,
+        "from_pretrained",
+        lambda *args, **kwargs: hf_config,
+    )
+    if model_class_name == "AutoModelForCausalLM":
+        monkeypatch.setattr(example_utils, "AutoModelForCausalLM", FakeAutoModelForCausalLM)
+    else:
+        monkeypatch.setattr(example_utils.transformers, model_class_name, FakeLlamaForCausalLM)
+    monkeypatch.setattr(example_utils, "is_nemotron_vl", lambda config: False)
+    monkeypatch.setattr(example_utils, "is_speculative", lambda config: False)
+    monkeypatch.setattr(example_utils, "init_empty_weights", lambda include_buffers: nullcontext())
+    monkeypatch.setattr(example_utils, "get_max_memory", lambda: {0: 1024})
+    monkeypatch.setattr(example_utils, "infer_auto_device_map", lambda model, max_memory: {"": 0})
+
+    model = example_utils.get_model("checkpoint", device="cpu", trust_remote_code=True)
+
+    assert isinstance(model, FakeModel)
+    assert calls["eval"]
+    if expected_config_dtype_kwarg == "torch_dtype":
+        assert calls["from_config"]["trust_remote_code"] is True
+    else:
+        assert "trust_remote_code" not in calls["from_config"]
+    assert calls["from_pretrained"]["trust_remote_code"] is True