Improve HF device handling

spacy_llm/models/hf/base.py

@@ -27,12 +27,46 @@ def __init__(
         inference_config (Dict[Any, Any]): HF config for model run.
         """
         self._name = name if self.hf_account in name else f"{self.hf_account}/{name}"
-        self._config_init, self._config_run = self.compile_default_configs()
+        default_cfg_init, default_cfg_run = self.compile_default_configs()
+        self._config_init, self._config_run = default_cfg_init, default_cfg_run
+
         if config_init:
             self._config_init = {**self._config_init, **config_init}
         if config_run:
             self._config_run = {**self._config_run, **config_run}
 
+        # `device` and `device_map` are conflicting arguments - ensure they aren't both set.
+        if config_init:
+            # Case 1: both device and device_map explicitly set by user.
+            if "device" in config_init and "device_map" in config_init:
+                warnings.warn(
+                    "`device` and `device_map` are conflicting arguments - don't set both. Dropping argument "
+                    "`device`."
+                )
+                self._config_init.pop("device")
+            # Case 2: we have a CUDA GPU (and hence device="cuda:0" by default), but device_map is set by user.
+            elif "device" in default_cfg_init and "device_map" in config_init:
+                self._config_init.pop("device")
+            # Case 3: we don't have a CUDA GPU (and hence "device_map=auto" by default), but device is set by user.
+            elif "device_map" in default_cfg_init and "device" in config_init:
+                self._config_init.pop("device_map")
+
+        # Fetch proper torch.dtype, if specified.
+        if (
+            has_torch
+            and "torch_dtype" in self._config_init
+            and self._config_init["torch_dtype"] != "auto"
+        ):
+            try:
+                self._config_init["torch_dtype"] = getattr(
+                    torch, self._config_init["torch_dtype"]
+                )
+            except AttributeError as ex:
+                raise ValueError(
+                    f"Invalid value {self._config_init['torch_dtype']} was specified for `torch_dtype`. "
+                    f"Double-check you specified a valid dtype."
+                ) from ex
+
         # Init HF model.
         HuggingFace.check_installation()
         self._check_model()
@@ -89,7 +123,7 @@ def compile_default_configs() -> Tuple[Dict[str, Any], Dict[str, Any]]:
         default_cfg_run: Dict[str, Any] = {}
 
         if has_torch:
-            default_cfg_init["torch_dtype"] = torch.bfloat16
+            default_cfg_init["torch_dtype"] = "bfloat16"
             if has_torch_cuda_gpu:
                 # this ensures it fails explicitely when GPU is not enabled or sufficient
                 default_cfg_init["device"] = "cuda:0"
@@ -106,6 +140,7 @@ def compile_default_configs() -> Tuple[Dict[str, Any], Dict[str, Any]]:
                     "Install CUDA to load and run the LLM on the GPU, or install 'accelerate' to dynamically "
                     "distribute the LLM on the CPU or even the hard disk. The latter may be slow."
                 )
+
         return default_cfg_init, default_cfg_run
 
     @abc.abstractmethod

spacy_llm/models/hf/falcon.py

@@ -19,7 +19,6 @@ def __init__(
         config_run: Optional[Dict[str, Any]],
     ):
         self._tokenizer: Optional["transformers.AutoTokenizer"] = None
-        self._device: Optional[str] = None
         super().__init__(name=name, config_init=config_init, config_run=config_run)
 
         assert isinstance(self._tokenizer, transformers.PreTrainedTokenizerBase)

spacy_llm/models/hf/mistral.py

@@ -1,4 +1,4 @@
-from typing import Any, Callable, Dict, Iterable, Optional, Tuple
+from typing import Any, Callable, Dict, Iterable, Optional
 
 from confection import SimpleFrozenDict
 
@@ -17,7 +17,6 @@ def __init__(
         config_run: Optional[Dict[str, Any]],
     ):
         self._tokenizer: Optional["transformers.AutoTokenizer"] = None
-        self._device: Optional[str] = None
         self._is_instruct = "instruct" in name
         super().__init__(name=name, config_init=config_init, config_run=config_run)
 
@@ -33,14 +32,15 @@ def __init__(
     def init_model(self) -> Any:
         self._tokenizer = transformers.AutoTokenizer.from_pretrained(self._name)
         init_cfg = self._config_init
+        device: Optional[str] = None
         if "device" in init_cfg:
-            self._device = init_cfg.pop("device")
+            device = init_cfg.pop("device")
 
         model = transformers.AutoModelForCausalLM.from_pretrained(
             self._name, **init_cfg, resume_download=True
         )
-        if self._device:
-            model.to(self._device)
+        if device:
+            model.to(device)
 
         return model
 
@@ -61,8 +61,7 @@ def __call__(self, prompts: Iterable[str]) -> Iterable[str]:  # type: ignore[ove
             ).input_ids
             for prompt in prompts
         ]
-        if self._device:
-            tokenized_input_ids = [tp.to(self._device) for tp in tokenized_input_ids]
+        tokenized_input_ids = [tp.to(self._model.device) for tp in tokenized_input_ids]
 
         return [
             self._tokenizer.decode(
@@ -74,14 +73,6 @@ def __call__(self, prompts: Iterable[str]) -> Iterable[str]:  # type: ignore[ove
             for tok_ii in tokenized_input_ids
         ]
 
-    @staticmethod
-    def compile_default_configs() -> Tuple[Dict[str, Any], Dict[str, Any]]:
-        default_cfg_init, default_cfg_run = HuggingFace.compile_default_configs()
-        return (
-            default_cfg_init,
-            default_cfg_run,
-        )
-
 
 @registry.llm_models("spacy.Mistral.v1")
 def mistral_hf(

spacy_llm/models/hf/openllama.py

@@ -2,7 +2,7 @@
 
 from confection import SimpleFrozenDict
 
-from ...compat import Literal, torch, transformers
+from ...compat import Literal, transformers
 from ...registry.util import registry
 from .base import HuggingFace
 
@@ -22,7 +22,6 @@ def __init__(
         config_run: Optional[Dict[str, Any]],
     ):
         self._tokenizer: Optional["transformers.AutoTokenizer"] = None
-        self._device: Optional[str] = None
         super().__init__(name=name, config_init=config_init, config_run=config_run)
 
     def init_model(self) -> "transformers.AutoModelForCausalLM":
@@ -32,14 +31,15 @@ def init_model(self) -> "transformers.AutoModelForCausalLM":
         # Initialize tokenizer and model.
         self._tokenizer = transformers.AutoTokenizer.from_pretrained(self._name)
         init_cfg = self._config_init
+        device: Optional[str] = None
         if "device" in init_cfg:
-            self._device = init_cfg.pop("device")
+            device = init_cfg.pop("device")
+
         model = transformers.AutoModelForCausalLM.from_pretrained(
             self._name, **init_cfg
         )
-
-        if self._device:
-            model.to(self._device)
+        if device:
+            model.to(device)
 
         return model
 
@@ -48,8 +48,9 @@ def __call__(self, prompts: Iterable[str]) -> Iterable[str]:  # type: ignore[ove
         tokenized_input_ids = [
             self._tokenizer(prompt, return_tensors="pt").input_ids for prompt in prompts
         ]
-        if self._device:
-            tokenized_input_ids = [tii.to(self._device) for tii in tokenized_input_ids]
+        tokenized_input_ids = [
+            tii.to(self._model.device) for tii in tokenized_input_ids
+        ]
 
         assert hasattr(self._model, "generate")
         return [
@@ -71,7 +72,7 @@ def compile_default_configs() -> Tuple[Dict[str, Any], Dict[str, Any]]:
         return (
             {
                 **default_cfg_init,
-                "torch_dtype": torch.float16,
+                "torch_dtype": "float16",
             },
             {**default_cfg_run, "max_new_tokens": 32},
         )

spacy_llm/models/hf/stablelm.py

@@ -42,7 +42,6 @@ def __init__(
     ):
         self._tokenizer: Optional["transformers.AutoTokenizer"] = None
         self._is_tuned = "tuned" in name
-        self._device: Optional[str] = None
         super().__init__(name=name, config_init=config_init, config_run=config_run)
 
     def init_model(self) -> "transformers.AutoModelForCausalLM":
@@ -51,14 +50,15 @@ def init_model(self) -> "transformers.AutoModelForCausalLM":
         """
         self._tokenizer = transformers.AutoTokenizer.from_pretrained(self._name)
         init_cfg = self._config_init
+        device: Optional[str] = None
         if "device" in init_cfg:
-            self._device = init_cfg.pop("device")
+            device = init_cfg.pop("device")
+
         model = transformers.AutoModelForCausalLM.from_pretrained(
             self._name, **init_cfg
         )
-
-        if self._device:
-            model.half().to(self._device)
+        if device:
+            model.half().to(device)
 
         return model
 
@@ -80,8 +80,7 @@ def __call__(self, prompts: Iterable[str]) -> Iterable[str]:  # type: ignore[ove
                 ]
             )
         ]
-        if self._device:
-            tokenized_input_ids = [tp.to(self._device) for tp in tokenized_input_ids]
+        tokenized_input_ids = [tp.to(self._model.device) for tp in tokenized_input_ids]
 
         assert hasattr(self._model, "generate")
         return [

spacy_llm/tests/models/test_hf.py

@@ -0,0 +1,78 @@
+from typing import Tuple
+
+import pytest
+import spacy
+from thinc.compat import has_torch_cuda_gpu
+
+from spacy_llm.compat import has_accelerate, torch
+
+_PIPE_CFG = {
+    "model": {
+        "@llm_models": "",
+        "name": "",
+    },
+    "task": {"@llm_tasks": "spacy.NoOp.v1"},
+    "save_io": True,
+}
+
+
+@pytest.mark.gpu
+@pytest.mark.skipif(not has_torch_cuda_gpu, reason="needs GPU & CUDA")
+@pytest.mark.parametrize(
+    "model", (("spacy.Dolly.v1", "dolly-v2-3b"), ("spacy.Llama2.v1", "Llama-2-7b-hf"))
+)
+def test_device_config_conflict(model: Tuple[str, str]):
+    """Test device configuration."""
+    nlp = spacy.blank("en")
+    model, name = model
+    cfg = {**_PIPE_CFG, **{"model": {"@llm_models": model, "name": name}}}
+
+    # Set device only.
+    cfg["model"]["config_init"] = {"device": "cpu"}  # type: ignore[index]
+    nlp.add_pipe("llm", name="llm1", config=cfg)
+
+    # Set device_map only.
+    cfg["model"]["config_init"] = {"device_map": "auto"}  # type: ignore[index]
+    if has_accelerate:
+        nlp.add_pipe("llm", name="llm2", config=cfg)
+    else:
+        with pytest.raises(ImportError, match="requires Accelerate"):
+            nlp.add_pipe("llm", name="llm2", config=cfg)
+
+    # Set device_map and device.
+    cfg["model"]["config_init"] = {"device_map": "auto", "device": "cpu"}  # type: ignore[index]
+    with pytest.warns(UserWarning, match="conflicting arguments"):
+        if has_accelerate:
+            nlp.add_pipe("llm", name="llm3", config=cfg)
+        else:
+            with pytest.raises(ImportError, match="requires Accelerate"):
+                nlp.add_pipe("llm", name="llm3", config=cfg)
+
+    torch.cuda.empty_cache()
+
+
+@pytest.mark.gpu
+@pytest.mark.skipif(not has_torch_cuda_gpu, reason="needs GPU & CUDA")
+def test_torch_dtype():
+    """Test torch_dtype setting."""
+    nlp = spacy.blank("en")
+    cfg = {
+        **_PIPE_CFG,
+        **{"model": {"@llm_models": "spacy.Dolly.v1", "name": "dolly-v2-3b"}},
+    }
+
+    # Should be converted to torch.float16.
+    cfg["model"]["config_init"] = {"torch_dtype": "float16"}  # type: ignore[index]
+    llm = nlp.add_pipe("llm", name="llm1", config=cfg)
+    assert llm._model._config_init["torch_dtype"] == torch.float16
+
+    # Should remain "auto".
+    cfg["model"]["config_init"] = {"torch_dtype": "auto"}  # type: ignore[index]
+    nlp.add_pipe("llm", name="llm2", config=cfg)
+
+    # Should fail - nonexistent dtype.
+    cfg["model"]["config_init"] = {"torch_dtype": "float999"}  # type: ignore[index]
+    with pytest.raises(ValueError, match="Invalid value float999"):
+        nlp.add_pipe("llm", name="llm3", config=cfg)
+
+    torch.cuda.empty_cache()

spacy_llm/tests/models/test_mistral.py

@@ -48,6 +48,7 @@ def test_init():
 
 
 @pytest.mark.gpu
+@pytest.mark.skip(reason="CI runner needs more GPU memory")
 @pytest.mark.skipif(not has_torch_cuda_gpu, reason="needs GPU & CUDA")
 def test_init_from_config():
     orig_config = Config().from_str(_NLP_CONFIG)