[draft] prime-rl integration - #8630

Closed

biswapanda wants to merge 22 commits into

mainfrom

bis/parity-tokenize-tcp

Commits on Apr 19, 2026

feat(rl): add vLLM weight lifecycle routes, TITO proxy, and tokenize for Prime-RL integration
biswapanda
committed
e95c4be
fix: engine route URL path (/engine/ not /engine_route/), pause/resume uses engine_client directly, NvExtResponse construction
biswapanda
committed
38a339d
fix: TITO token_data bypass, reload_weights API, strip unsupported fields
biswapanda
committed
bdddd2b
--wip-- [skip ci]
biswapanda
committed
3b73f24
fix: rl_admin token_ids injection + publisher crash + error codes
biswapanda
committed
cc0ee09
fix: token_ids injection alignment for byte-level tokens
biswapanda
committed
61da199
fix: pre-tokenize prompts with Python HF tokenizer via nvext.token_data
biswapanda
committed
8ecc798
feat: add /tokenize and /detokenize endpoints (cherry-picked from PR #7699)
biswapanda
committed
e2f385d
refactor: rl_admin uses Rust /tokenize, removes Python HF tokenizer dep
biswapanda
committed
fb4c390
rl: wire completion_token_ids, /v1/rl/* router, /v1/tokenize prefix

bispnv
authored and
biswapanda
committed
03d58a9
rl: add GET /v1/rl/health for Prime-RL admin client health check

bispnv
authored and
biswapanda
committed
8cebd0f
rl: auto-enable token_ids when DYN_ENABLE_RL, promote prompt_token_ids + choice.token_ids

bispnv
authored and
biswapanda
committed
70f8457
add draft api doc

bispnv
authored and
biswapanda
committed
d9cd90d
rl: add /v1/rl/load_lora_adapter and /v1/rl/unload_lora_adapter
biswapanda
committed
60e52ee
deps: bump default VLLM_VER to 0.19.1
biswapanda
committed
f1d18d7
deps: install prime-rl plugin via pip to apply vllm.general_plugins patches
biswapanda
committed
e8ec876
deps: drop --ignore-requires-python (uv pip does not support it)
biswapanda
committed
73d581b
fix: adapt tokenize endpoint + default convert_ids_to_tokens to DecodeResult
biswapanda
committed
d837fbd

Navigation Menu

Uh oh!

[draft] prime-rl integration - #8630

[draft] prime-rl integration #8630
biswapanda wants to merge 22 commits into
mainfrom
bis/parity-tokenize-tcp

Commits on Apr 19, 2026

feat(rl): add vLLM weight lifecycle routes, TITO proxy, and tokenize for Prime-RL integration

fix: engine route URL path (/engine/ not /engine_route/), pause/resume uses engine_client directly, NvExtResponse construction

fix: TITO token_data bypass, reload_weights API, strip unsupported fields

--wip-- [skip ci]

fix: rl_admin token_ids injection + publisher crash + error codes

fix: token_ids injection alignment for byte-level tokens

fix: pre-tokenize prompts with Python HF tokenizer via nvext.token_data

feat: add /tokenize and /detokenize endpoints (cherry-picked from PR #7699)

refactor: rl_admin uses Rust /tokenize, removes Python HF tokenizer dep

rl: wire completion_token_ids, /v1/rl/* router, /v1/tokenize prefix

rl: add GET /v1/rl/health for Prime-RL admin client health check

rl: auto-enable token_ids when DYN_ENABLE_RL, promote prompt_token_ids + choice.token_ids

add draft api doc

rl: add /v1/rl/load_lora_adapter and /v1/rl/unload_lora_adapter

deps: bump default VLLM_VER to 0.19.1

deps: install prime-rl plugin via pip to apply vllm.general_plugins patches

deps: drop --ignore-requires-python (uv pip does not support it)

fix: adapt tokenize endpoint + default convert_ids_to_tokens to DecodeResult

Commits on Apr 23, 2026

fix(rl): use nvext.token_data as prompt_token_ids for TITO requests

chore(rl): remove rl_admin Python service (superseded by Rust frontend)

fix(rl): LoRA hot-swap state corruption, silent data-loss, and tokenizer asymmetry

fix: whitelist cache_salt in validate_no_unsupported_fields