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 15, 18:30 UTC
Revert "[Bugfix] Limit profiling run sequence length by max_model_len (#14785) (#14892)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Mar 17, 2025 at 00:13 UTC
f6137adbcbbdea8b5023a66480de921b558bef83
5 files changed
-9
vllm/inputs/registry.py
-5
@@ -330,11 +330,6 @@ class InputRegistry:
330
from vllm.multimodal import MultiModalKwargs
331
from vllm.multimodal.profiling import MultiModalProfiler
332
333
- if seq_len > model_config.max_model_len:
334
- raise AssertionError(
335
- f"Profiling attempted with sequence length ({seq_len}) "
336
- f"greater than model length ({model_config.max_model_len})")
337
-
333
if mm_registry.has_processor(model_config):
334
tokenizer = cached_tokenizer_from_config(model_config)
335
processor = mm_registry.create_processor(model_config,
vllm/worker/enc_dec_model_runner.py
-1
@@ -281,7 +281,6 @@ class EncoderDecoderModelRunner(GPUModelRunnerBase[EncoderDecoderModelInput]):
281
for group_id in range(max_num_seqs):
282
seq_len = (max_num_batched_tokens // max_num_seqs +
283
(group_id < max_num_batched_tokens % max_num_seqs))
284
- seq_len = min(seq_len, self.model_config.max_model_len)
284
batch_size += seq_len
285
286
decoder_dummy_data = self.input_registry \
vllm/worker/model_runner.py
-1
@@ -1302,7 +1302,6 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
1302
for group_id in range(max_num_seqs):
1303
seq_len = (max_num_batched_tokens // max_num_seqs +
1304
(group_id < max_num_batched_tokens % max_num_seqs))
1305
- seq_len = min(seq_len, self.model_config.max_model_len)
1305
batch_size += seq_len
1306
1307
dummy_data = self.input_registry \
vllm/worker/openvino_model_runner.py
-1
@@ -148,7 +148,6 @@ class OpenVINOModelRunner(ModelRunnerBase):
148
seq_len = min(
149
seq_data.get_len(),
150
computed_len + seq_group_metadata.token_chunk_size,
151
- self.model_config.max_model_len,
151
)
152
if is_prompt:
153
tokens = seq_data.get_token_ids()[computed_len:seq_len]
vllm/worker/xpu_model_runner.py
-1
@@ -466,7 +466,6 @@ class XPUModelRunner(ModelRunnerBase[ModelInputForXPUWithSamplingMetadata]):
466
for group_id in range(max_num_seqs):
467
seq_len = (max_num_batched_tokens // max_num_seqs +
468
(group_id < max_num_batched_tokens % max_num_seqs))
469
- seq_len = min(seq_len, self.model_config.max_model_len)
469
batch_size += seq_len
470
471
dummy_data = self.input_registry \