Files
Xiaoze Fan 08d728d5e3 feat(mm): serve image input on the Qwen families (#454)
* feat(mm): serve image input on the Qwen VL families

- frontend: OpenAI image_url, Anthropic image blocks and Responses
  input_image (http(s), base64, gated file://) become MMItems in the
  tokenizer workers; image spans carry content pseudo ids so the radix
  cache keys them by content; mrope positions are precomputed per request
- registry: EncoderSpec per family (Qwen3-VL, Qwen3.6, Qwen3.8-Flash-Next
  register a vision encoder serving image) and ModelSpec.mm_processor;
  EngineConfig.active_encoders decides what a process builds,
  served_modalities what it accepts; --text-model-only, --mm-disable,
  --mm-encoder-weights, --mm-embed-cache-device, --mm-max-pixels
- engine and scheduler: per-request encoder-cache row ownership, chunk
  planning on the batch (jobs, gather rows, scatter rows), items encoded
  right before the LM forward and gathered into batch.mm_embeds
- models: the Qwen VL vision tower on the text tower's TP layers with
  DeepStack, block weights streamed from pinned host banks by default,
  merger outputs parked on the host between blocks; SupportsMultimodal
  (encode, place_encoder_weights) and a shared embed_input_ids
- server: image inputs rejected with a client-facing reason when no
  vision encoder is served; GET /v1/stats reports model.input_modalities

* feat(mm): replace the pixel cap with a token budget and processor kwargs

- --image-min-tokens / --image-max-tokens replace --mm-max-pixels; MMProcessor.get_mm_processor_kwargs turns the MultimodalConfig into the processor call's arguments, the default passes mm.processor_kwargs through and Qwen VL converts the budget to pixel areas in size.shortest_edge / longest_edge
- --mm-processor-kwargs hands a JSON object to the checkpoint's image processor call, applied after the token budget
- MMProcessor is an ABC that owns the model path, the config and the lazily loaded image processor; get_mm_processor returns None when --mm-disable leaves no encoder to serve
- the frontend tokenizer, the tokenizer worker and the offline LLM build one processor each and hand it to TokenizeManager, so the tokenize message carries no per-request pixel cap
2026-09-13 15:39:33 -07:00

34 lines
1.1 KiB
Python

"""Multimodal input handling: pad-id scheme, per-family processors, media fetch, embedding cache."""
from __future__ import annotations
import torch
# image spans become content pseudo ids above every real token id: the radix cache keys them by content and the model finds image rows by id >= MM_PAD_SHIFT_VALUE
# they are clamped back into the vocab before the embedding lookup; those rows are overwritten anyway
MM_PAD_SHIFT_VALUE = 1_000_000
def restore_placeholder(input_ids: torch.Tensor, token_id: int) -> torch.Tensor:
"""input_ids with every content pad id put back to the family's placeholder token."""
return input_ids.masked_fill(input_ids >= MM_PAD_SHIFT_VALUE, token_id)
def mm_pad_value(item_hash: int) -> int:
return MM_PAD_SHIFT_VALUE + item_hash % (1 << 30)
def check_mm_pad_shift(vocab_size: int) -> None:
if vocab_size > MM_PAD_SHIFT_VALUE:
raise ValueError(
f"vocab_size ({vocab_size}) exceeds MM_PAD_SHIFT_VALUE ({MM_PAD_SHIFT_VALUE}); "
"image pseudo-ids would alias real token ids"
)
__all__ = [
"MM_PAD_SHIFT_VALUE",
"check_mm_pad_shift",
"mm_pad_value",
]