Skip to content

Independent legolink segspans cache#6

Open
omerpaz95 wants to merge 32 commits into
mainfrom
independent_legolink_segspans_cache
Open

Independent legolink segspans cache#6
omerpaz95 wants to merge 32 commits into
mainfrom
independent_legolink_segspans_cache

Conversation

@omerpaz95

Copy link
Copy Markdown
Owner

I still have to review this thoroughly

David Ben-David and others added 30 commits January 21, 2026 17:00
Signed-off-by: David Ben-David <davidb@pliops.com>
…ed-token gaps

Signed-off-by: David Ben-David <davidb@pliops.com>
Signed-off-by: Kfir Wolfson <kfirw@pliops.com>
Signed-off-by: Omer Paz <omerpaz95@gmail.com>
Signed-off-by: omerpaz95 <omerpaz95@gmail.com>
Signed-off-by: omerpaz95 <omerpaz95@gmail.com>
Signed-off-by: omerpaz95 <omerpaz95@gmail.com>
…stic.

Signed-off-by: omerpaz95 <omerpaz95@gmail.com>
Signed-off-by: omerpaz95 <omerpaz95@gmail.com>
Signed-off-by: omerpaz95 <omerpaz95@gmail.com>
Make block size, gap policy, gap length, and pad token configurable
via VLLM_V1_SPANS_* environment variables so they can be set in
Docker deployments without CLI args. Also register the missing
--gap-policy-name and --gap-policy-config CLI arguments that existed
as EngineArgs fields but were not exposed to argparse.

New env vars:
- VLLM_V1_SPANS_PAD_TOKEN
- VLLM_V1_SPANS_BLOCK_SIZE
- VLLM_V1_SPANS_GAP_POLICY_ENABLE
- VLLM_V1_SPANS_GAP_LENGTH
Add env var configuration for span/gap-policy parameters
…c inconsistency between vLLM regular path and spans path, matching the triton attention kernel's fused RoPE to the RoPE CUDA kernel
Signed-off-by: omerpaz95 <omerpaz95@gmail.com>
Signed-off-by: omerpaz95 <omerpaz95@gmail.com>
Signed-off-by: omerpaz95 <omerpaz95@gmail.com>
…ries

   Span boundaries are no longer marked by special tokens (SPAN_TOKEN_PLUS,
   SPAN_TOKEN_CROSS) embedded in the token stream. Instead, span_starts and
   cross_span_starts are passed via SamplingParams.extra_args and read by
   Request and SpanAwareGapPolicy directly.

Signed-off-by: omerpaz95 <omerpaz95@gmail.com>
Remove span marker tokens from prompt, use extra_args for span bounda…
@omerpaz95 omerpaz95 self-assigned this Apr 23, 2026
Signed-off-by: omerpaz95 <omerpaz95@gmail.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants