@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 0790f07695

Fork 0

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 10, 14:00 UTC

[Misc] Improve error messages for unsupported types and parameters (#30593)

Signed-off-by: BlankR <hjyblanche@gmail.com> Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>

BlankR committed Jan 7, 2026 at 01:00 UTC 0790f07695c72fe203e95f6d8c8ff15c8003e754

11 files changed +40 -16

benchmarks/cutlass_benchmarks/sparse_benchmarks.py

+3 -1

@@ -343,7 +343,9 @@ def bench(

343 return bench_int8(dtype, m, k, n, label, sub_label)

344 if dtype == torch.float8_e4m3fn:

345 return bench_fp8(dtype, m, k, n, label, sub_label)

346 - raise ValueError("unsupported type")

346 + raise ValueError(

347 + f"Unsupported dtype {dtype}: should be one of torch.int8, torch.float8_e4m3fn."

348 + )

349

350

351 # runner

vllm/attention/ops/chunked_prefill_paged_decode.py

+4 -1

@@ -292,7 +292,10 @@ def chunked_prefill_paged_decode(

292 elif kv_cache_dtype == "fp8_e5m2":

293 target_dtype = torch.float8_e5m2

294 else:

295 - raise ValueError("Unsupported FP8 dtype:", kv_cache_dtype)

295 + raise ValueError(

296 + f"Unsupported FP8 kv_cache_dtype {kv_cache_dtype}: "

297 + f"should be one of 'fp8', 'fp8_e4m3', 'fp8_e5m2'."

298 + )

299

300 key_cache = key_cache.view(target_dtype)

301 value_cache = value_cache.view(target_dtype)

vllm/config/lora.py

+1 -1

@@ -90,7 +90,7 @@ class LoRAConfig:

90 elif self.max_cpu_loras < self.max_loras:

91 raise ValueError(

92 f"max_cpu_loras ({self.max_cpu_loras}) must be >= "

93 - f"max_loras ({self.max_loras})"

93 + f"max_loras ({self.max_loras})."

94 )

96 return self

vllm/distributed/device_communicators/pynccl_wrapper.py

+4 -1

@@ -92,7 +92,10 @@ class ncclDataTypeEnum:

92 return cls.ncclFloat64

93 if dtype == torch.bfloat16:

94 return cls.ncclBfloat16

95 - raise ValueError(f"Unsupported dtype: {dtype}")

95 + raise ValueError(

96 + f"Unsupported dtype {dtype}: should be one of "

97 + f"int8, uint8, int32, int64, float16, float32, float64, bfloat16."

98 + )

100

101 ncclRedOp_t = ctypes.c_int

vllm/distributed/kv_transfer/kv_connector/v1/lmcache_integration/vllm_v1_adapter.py

+4 -1

@@ -233,7 +233,10 @@ class RequestTracker:

233 elif isinstance(new_block_ids, list):

234 pass

235 else:

236 - raise ValueError(f"Unsupported new_block_ids type {type(new_block_ids)}")

236 + raise ValueError(

237 + f"Unsupported new_block_ids type {type(new_block_ids)}: "

238 + f"should be None[list[int], ...], tuple or list[int]."

239 + )

240 self.allocated_block_ids.extend(new_block_ids)

241

242 # When a request is scheduled again, and the number of new tokens

vllm/model_executor/layers/quantization/auto_round.py

+6 -6

@@ -56,22 +56,22 @@ class AutoRoundConfig(QuantizationConfig):

56 if weight_bits not in self.SUPPORTED_BITS:

57 raise ValueError(

58 f"Unsupported weight_bits: {weight_bits}, "

59 - f"currently only support {self.SUPPORTED_BITS}"

59 + f"currently only support {self.SUPPORTED_BITS}."

60 )

61 if data_type not in self.SUPPORTED_DTYPES:

62 raise ValueError(

63 - f"Unsupported data_type: {data_type},"

64 - f" currently only support {self.SUPPORTED_DTYPES}"

63 + f"Unsupported data_type: {data_type}, "

64 + f"currently only support {self.SUPPORTED_DTYPES}."

65 )

66 if packing_format not in self.SUPPORTED_FORMATS:

67 raise ValueError(

68 f"Unsupported packing_format: {packing_format}, "

69 - f"currently only support {self.SUPPORTED_FORMATS}"

69 + f"currently only support {self.SUPPORTED_FORMATS}."

70 )

71 if backend not in self.SUPPORTED_BACKENDS:

72 raise ValueError(

73 - f"Unsupported backend: {backend}, "

74 - f"currently only support {self.SUPPORTED_BACKENDS}"

73 + f"Unsupported backend: {backend}, "

74 + f"currently only support {self.SUPPORTED_BACKENDS}."

75 )

77 self.weight_bits = weight_bits

vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_fp8.py

+4 -1

@@ -158,7 +158,10 @@ class CompressedTensorsW8A8Fp8(CompressedTensorsScheme):

158 input_scale = None

159

160 else:

161 - raise ValueError(f"Unknown quantization strategy {self.strategy}")

161 + raise ValueError(

162 + f"Unknown quantization strategy {self.strategy}: "

163 + f"should be one of {list(QuantizationStrategy)}"

164 + )

165

166 # required by torch.compile to be torch.nn.Parameter

167 layer.weight = Parameter(weight.data, requires_grad=False)

vllm/model_executor/layers/quantization/mxfp4.py

+4 -1

@@ -783,7 +783,10 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):

783 layer.w13_weight = w13_weight

784 layer.w2_weight = w2_weight

785 else:

786 - raise ValueError(f"Unsupported backend: {self.mxfp4_backend}")

786 + raise ValueError(

787 + f"Unsupported mxfp4_backend: {self.mxfp4_backend}: "

788 + f"should be one of: {list(Mxfp4Backend)}."

789 + )

790

791 def get_fused_moe_quant_config(

792 self, layer: torch.nn.Module

vllm/model_executor/models/ernie45_vl.py

+5 -1

@@ -599,7 +599,11 @@ def smart_resize(

599 w_bar = ceil_by_factor(width * beta, factor)

600

601 if min_pixels > h_bar * w_bar or h_bar * w_bar > max_pixels:

602 - raise ValueError(f"encounter invalid h_bar: {h_bar}, w_bar: {w_bar}")

602 + raise ValueError(

603 + f"Invalid h_bar={h_bar}, w_bar={w_bar}: "

604 + f"h_bar * w_bar must be >= min_pixels ({min_pixels}) "

605 + f"and <= max_pixels ({max_pixels})."

606 + )

607

608 return h_bar, w_bar

609

vllm/model_executor/models/granite_speech.py

+3 -1

@@ -348,7 +348,9 @@ class GraniteSpeechConformerAttention(nn.Module):

348

349 if self.context_size <= 0 or self.context_size > self.max_pos_emb:

350 raise ValueError(

351 - "Context size is either less than 0 or exceeds the max_pos_emb"

351 + f"Context size should be > 0 and "

352 + f"<= max_pos_emb ({self.max_pos_emb}), "

353 + f"got {self.context_size}."

354 )

355

356 def forward(

vllm/model_executor/models/minimax_text_01.py

+2 -1

@@ -332,7 +332,8 @@ class MiniMaxText01DecoderLayer(nn.Module):

332 )

333 else:

334 raise ValueError(

335 - f"Unsupported attention type: {self.config.attention_type}"

335 + f"Unsupported attention_type {self.config.attention_type}: "

336 + f"should be 0 (linear) or 1 (full)."

337 )

338

339 if expert_num == 1: