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 14, 05:30 UTC
[Doc] Fix cross-reference warnings (#25058)
Signed-off-by: Punit Vara <punitvara@gmail.com> Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Punitvara committed
Sep 18, 2025 at 14:35 UTC
05b044e698bb3c151871d94b64fabd87188de9ef
6 files changed
+18
-14
vllm/benchmarks/datasets.py
+2
-1
@@ -171,7 +171,8 @@ class BenchmarkDataset(ABC):
171
If `None`, LoRA is not used.
172
173
Returns:
174
- A new [LoRARequest][] (or `None` if not applicable).
174
+ A new [`LoRARequest`][vllm.lora.request.LoRARequest]
175
+ (or `None` if not applicable).
176
"""
177
if max_loras is None or lora_path is None:
178
return None
vllm/distributed/device_communicators/shm_object_storage.py
+4
-4
@@ -30,7 +30,7 @@ class SingleWriterShmRingBuffer:
30
- Maintains metadata for each allocated buffer chunk in the writer process
31
- Supports custom "is_free_fn" functions to determine when buffers can be
32
reused
33
- - Each buffer chunk contains: [4-byte id][4-byte size][actual_data]
33
+ - Each buffer chunk contains: `[4-byte id][4-byte size][actual_data]`
34
35
Key Concepts:
36
- monotonic_id_start/end: Track the range of active buffer IDs
@@ -99,7 +99,7 @@ class SingleWriterShmRingBuffer:
99
- Writer handles garbage collection (free_buf) based on reader feedback
100
101
Memory Layout per Buffer Chunk:
102
- [4-byte monotonic_id][4-byte chunk_size][actual_data...]
102
+ `[4-byte monotonic_id][4-byte chunk_size][actual_data...]`
103
^metadata_start ^data_start
104
105
The monotonic_id ensures data integrity - readers can verify they're
@@ -185,7 +185,7 @@ class SingleWriterShmRingBuffer:
185
'''
186
Allocate a buffer `MD_SIZE` + `size` bytes in the shared memory.
187
Memory layout:
188
- [4-byte monotonic_id][4-byte size][buffer data...]
188
+ `[4-byte monotonic_id][4-byte size][buffer data...]`
189
'''
190
assert self.is_writer, "Only the writer can allocate buffers."
191
assert size > 0, "Size must be greater than 0"
@@ -413,7 +413,7 @@ class SingleWriterShmObjectStorage:
413
allocation
414
415
Memory Layout per Object:
416
- [4-byte reference_count][metadata_size][serialized_object_data]
416
+ `[4-byte reference_count][metadata_size][serialized_object_data]`
417
418
Thread Safety:
419
- Writer operations (put, clear) are single-threaded by design
vllm/model_executor/layers/mamba/ops/causal_conv1d.py
+7
-5
@@ -927,11 +927,13 @@ def causal_conv1d_update(
927
validate_data=False,
928
):
929
"""
930
- x: (batch, dim) or (batch, dim, seqlen) or (num_tokens, dim)
931
- [shape=2: single token prediction]
932
- [shape=3: single or multiple tokens prediction]
933
- [shape=2 with num_tokens: continuous batching, where num_tokens is the
934
- total tokens of all sequences in that batch]
930
+ x: Input tensor which can take the following shapes:
931
+
932
+ - `[batch, dim]` - single token prediction
933
+ - `[batch, dim, seqlen]` - single or multiple tokens prediction
934
+ - `[num_tokens, dim]` - continuous batching, where num_tokens is
935
+ the total tokens of all sequences in that batch
936
+
937
conv_state: (..., dim, state_len), where state_len >= width - 1
938
weight: (dim, width)
939
bias: (dim,)
vllm/model_executor/models/mistral3.py
+1
-1
@@ -583,7 +583,7 @@ class Mistral3ForConditionalGeneration(nn.Module, SupportsLoRA,
583
inputs_embeds: Optional tensor of input embeddings.
584
585
Info:
586
- [Mistral3ImagePixelInputs][]
586
+ [`Mistral3ImagePixelInputs`][vllm.model_executor.models.mistral3.Mistral3ImagePixelInputs]
587
"""
588
if intermediate_tensors is not None:
589
inputs_embeds = None
vllm/multimodal/profiling.py
+1
-1
@@ -301,7 +301,7 @@ class MultiModalProfiler(Generic[_I]):
301
Returns the maximum length of the multimodal (image placeholders+text)
302
tokens, including any break/text tokens in-between image embeddings.
303
304
- <im_start> [IMG] [IMG] [IMG] <row_break> [IMG] [IMG] [IMG] <im_end>
304
+ `<im_start> [IMG] [IMG] [IMG] <row_break> [IMG] [IMG] [IMG] <im_end>`
305
Returns 9, even when the number of image embeddings is 6.
306
307
This is important to take into account when profiling and
vllm/v1/core/kv_cache_manager.py
+3
-2
@@ -24,8 +24,9 @@ class KVCacheBlocks:
24
"""
25
blocks: tuple[list[KVCacheBlock], ...]
26
"""
27
- blocks[i][j] refers to the i-th kv_cache_group and the j-th block of tokens.
28
- We don't use block of tokens as the outer dimension because it assumes all
27
+ `blocks[i][j]` refers to the i-th kv_cache_group
28
+ and the j-th block of tokens.We don't use block of
29
+ tokens as the outer dimension because it assumes all
30
kv_cache_groups have the same number of blocks, which is true for now but
31
will be broken if we want to give different block_size to different
32
kv_cache_groups in the future.