A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 03:15 UTC
[mypy] Pass type checking in vllm/inputs (#11680)
Signed-off-by: Tobias Pitters <tobias.pitters@gmail.com>
Tobias Pitters committed
Jan 2, 2025 at 17:18 UTC
b6087a6beead9165f4c77ceba592b3651bb37de9
4 files changed
+16
-14
tools/mypy.sh
+1
@@ -23,6 +23,7 @@ run_mypy vllm/compilation
23
run_mypy vllm/distributed
24
run_mypy vllm/engine
25
run_mypy vllm/executor
26
+run_mypy vllm/inputs
27
run_mypy vllm/lora
28
run_mypy vllm/model_executor
29
run_mypy vllm/plugins
vllm/inputs/data.py
+11
-10
@@ -250,7 +250,7 @@ class SingletonInputsAdapter:
250
if inputs["type"] == "token" or inputs["type"] == "multimodal":
251
return inputs.get("prompt")
252
253
- assert_never(inputs)
253
+ assert_never(inputs) # type: ignore[arg-type]
254
255
@cached_property
256
def prompt_token_ids(self) -> List[int]:
@@ -259,7 +259,7 @@ class SingletonInputsAdapter:
259
if inputs["type"] == "token" or inputs["type"] == "multimodal":
260
return inputs.get("prompt_token_ids", [])
261
262
- assert_never(inputs)
262
+ assert_never(inputs) # type: ignore[arg-type]
263
264
@cached_property
265
def token_type_ids(self) -> List[int]:
@@ -268,7 +268,7 @@ class SingletonInputsAdapter:
268
if inputs["type"] == "token" or inputs["type"] == "multimodal":
269
return inputs.get("token_type_ids", [])
270
271
- assert_never(inputs)
271
+ assert_never(inputs) # type: ignore[arg-type]
272
273
@cached_property
274
def prompt_embeds(self) -> Optional[torch.Tensor]:
@@ -277,7 +277,7 @@ class SingletonInputsAdapter:
277
if inputs["type"] == "token" or inputs["type"] == "multimodal":
278
return None
279
280
- assert_never(inputs)
280
+ assert_never(inputs) # type: ignore[arg-type]
281
282
@cached_property
283
def multi_modal_data(self) -> "MultiModalDataDict":
@@ -289,7 +289,7 @@ class SingletonInputsAdapter:
289
if inputs["type"] == "multimodal":
290
return inputs.get("mm_kwargs", {})
291
292
- assert_never(inputs)
292
+ assert_never(inputs) # type: ignore[arg-type]
293
294
@cached_property
295
def multi_modal_inputs(self) -> Union[Dict, "MultiModalKwargs"]:
@@ -301,7 +301,7 @@ class SingletonInputsAdapter:
301
if inputs["type"] == "multimodal":
302
return inputs.get("mm_kwargs", {})
303
304
- assert_never(inputs)
304
+ assert_never(inputs) # type: ignore[arg-type]
305
306
@cached_property
307
def multi_modal_hashes(self) -> List[str]:
@@ -311,9 +311,10 @@ class SingletonInputsAdapter:
311
return inputs.get("multi_modal_hashes", [])
312
313
if inputs["type"] == "multimodal":
314
- return inputs.get("mm_hashes", [])
314
+ # only the case when we use MultiModalInputsV2
315
+ return inputs.get("mm_hashes", []) # type: ignore[return-value]
316
316
- assert_never(inputs)
317
+ assert_never(inputs) # type: ignore[arg-type]
318
319
@cached_property
320
def multi_modal_placeholders(self) -> "MultiModalPlaceholderDict":
@@ -325,7 +326,7 @@ class SingletonInputsAdapter:
326
if inputs["type"] == "multimodal":
327
return inputs.get("mm_placeholders", {})
328
328
- assert_never(inputs)
329
+ assert_never(inputs) # type: ignore[arg-type]
330
331
@cached_property
332
def mm_processor_kwargs(self) -> Dict[str, Any]:
@@ -337,7 +338,7 @@ class SingletonInputsAdapter:
338
if inputs["type"] == "multimodal":
339
return {}
340
340
- assert_never(inputs)
341
+ assert_never(inputs) # type: ignore[arg-type]
342
343
344
ProcessorInputs = Union[DecoderOnlyInputs, EncoderDecoderInputs]
vllm/inputs/preprocess.py
+3
-3
@@ -436,7 +436,7 @@ class InputPreprocessor:
436
or encoder_inputs["type"] == "multimodal"):
437
pass
438
else:
439
- assert_never(encoder_inputs)
439
+ assert_never(encoder_inputs) # type: ignore[arg-type]
440
441
if decoder_inputs is None:
442
dec_token_ids = self._prepare_decoder_input_ids_for_generation(
@@ -452,7 +452,7 @@ class InputPreprocessor:
452
raise ValueError("Multi-modal decoder inputs of encoder-"
453
"decoder models are not supported yet")
454
else:
455
- assert_never(encoder_inputs)
455
+ assert_never(encoder_inputs) # type: ignore[arg-type]
456
457
return EncoderDecoderInputs(
458
encoder=encoder_inputs,
@@ -569,7 +569,7 @@ class InputPreprocessor:
569
prompt_adapter_request=prompt_adapter_request,
570
)
571
else:
572
- assert_never(prompt_inputs)
572
+ assert_never(prompt_inputs) # type: ignore[arg-type]
573
574
return prompt_inputs
575
vllm/inputs/registry.py
+1
-1
@@ -419,7 +419,7 @@ class InputRegistry:
419
# Be more strict in V2
420
assert "mm_kwargs" in inputs
421
else:
422
- assert_never(inputs["type"])
422
+ assert_never(inputs["type"]) # type: ignore[arg-type]
423
424
def process_input(self, model_config: "ModelConfig",
425
inputs: ProcessorInputs) -> ProcessorInputs: