Commit Graph
33 Commits
Author SHA1 Message Date
84523d6785 [Spec] Reuse K3 auxiliary outputs across decode CUDA graph sizes (#40159)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: hnyls2002 <lsyincs@gmail.com>
2026-09-29 14:24:51 -07:00
Liangsheng Yin f35ec18d8b [Test] Remove more unit tests that mirror implementation or never run in CI (#41297) 2026-09-25 19:22:14 -07:00
Liangsheng Yin cc6e8b0f25 [Fix] Derive per-runner hybrid SWA layer ids on ModelLayerInfo instead of mutating ModelConfig (#40983) 2026-09-23 19:52:55 -07:00
Cheng Wan 65be3fa71a A runner and the objects it builds freeze the placement they describe (#40341) 2026-09-21 12:24:17 -07:00
cctry 7a6191c4b9 Preallocate HiCache MHA staging before post-capture KV sizing (#40256) 2026-09-21 10:44:29 -07:00
Cheng Wan 3a5f52e144 Record a process's placement at publish, not at group build (#40071) 2026-09-18 23:52:37 -07:00
metamergebotandcctry 6a9c7001d3 [Logprob] Borrow graph-pool memory for input logprob logits construction (#40007)
Co-authored-by: cctry <csycfl@gmail.com>
2026-09-18 13:18:01 -07:00
4da5599e93 Fix first-token metadata and reused attention-layer indexing (#39328)
Co-authored-by: Jinghui Zhang <jinghui@meta.com>
Co-authored-by: Lucia Fang <116399278+luccafong@users.noreply.github.com>
2026-09-15 15:22:26 -07:00
cctryandcctry 206034e520 Keep graph-pool borrows on their allocation stream (#39180)
Co-authored-by: cctry <17473714+cctry@users.noreply.github.com>
2026-09-12 19:26:09 -07:00
cctry 203d7e812c Fix KV-canary workspace accounting after graph capture (#38596) 2026-09-10 16:10:12 -07:00
Cheng Wan aaf9a95763 [Config] Round 6.5: a namespace declares what it derives, next to what it derives it from (#38113)
Fifth of five; stacked on #38049. The split gave every namespace a file, but
only for the half an operator types. This is the other half.

## The parallel quotients are declared, not written out

`attn_tp_size` and its five siblings were sixty lines of near-identical
properties in the runtime context, a file away from the leaves they are
quotients of, so reading `parallel.py` told you what you could set and nothing
about what that decides.

They are declared in `Parallel` now, in the same class as those leaves. They
carry no annotation, so they are not dataclass fields and
`collect_input_fields` never puts them on the record -- the same mechanism that
already keeps `_NS_PATH` off it. That is the right exclusion: a quotient has no
operator input to preserve, and the record is what crosses a process boundary,
where a stamped width is one an elastic scale-up will not refresh.

## A quotient is a value in the bag, like every other derived one

`_derived_width` answered from a stamp or, failing that, a live process group.
The group read could never disagree with the stamp:

- `initialize_model_parallel` stamps all six as its last statement,
  unconditionally;
- an elastic scale-up restamps `attn_dp_size` through
  `update_dp_attention_post_scale` -- the comment claiming it does *not* was
  wrong;
- no hardware backend builds groups of its own;
- `multimodal_gen`, which has its own `initialize_model_parallel` and does not
  stamp, never reads a quotient.

So a built group was always already stamped, and the group read goes -- and with
it the last reason for a quotient to be resolved on every read.

Every input to `derive_parallel_widths` is a record field. `dcp_enabled` is
`decode_context_parallel_size > 1`, not a fact about a built group; it was
spelled `_DCP is not None`, which is a longer way to say the same thing. So the
six are fixed once the configuration is fixed -- the same test every other
`Derived(fn=...)` in this PR passes. They are declared the same way and computed
the same way: once, at publish, into ordinary bag leaves.

What remains is override -> stamp -> published leaf. The stamp stays above the
leaf because an elastic scale-up restamps `attn_dp_size`; the override stays on
top because that is how a test names a width.

## One answer for the config-derived predicates

`enable_mamba_extra_buffer` and its lazy variant, `is_ep_joiner`,
`is_ep_scale_joiner`, `is_startup_weight_load_overlap`: each existed as a
`ServerArgs` member for the resolution pipeline and, for most of them, again as
a `runtime_context` function for readers after publish. Three places to keep
saying the same thing.

A `Derived(fn=...)` is a pure function of the published configuration, so
`publish` computes it once and stores it as an ordinary bag leaf -- a plain
attribute load, which is what a read inside compiled model code needs. The
function is handed the whole resolved config rather than the bag it lands in,
because a derivation is free to span namespaces and the mamba one does: it
reads `memory.disable_radix_cache` alongside its own `exec.mamba` strategy,
which is why it could never have been a method on either bag.

The pre-publish helpers stay -- resolution needs the predicate before there is
a bag to read -- and three readers keep them, because they run before their own
process publishes: `initialize_dp_attention`, which the weight-cache daemon
calls while building its groups thirty lines before its `publish`, and
`PortArgs.init_new`, a factory handed the record that already reads eighteen
other fields off it.

## Notes for a reviewer

**Overriding a leaf does not move its quotient.** `override(tp_size=2)` leaves
`attn_tp_size` where the published config put it, because nothing is recomputed
on read. A test states a topology by publishing a config -- which is what a
real process does -- or by naming the width it wants, `override(attn_tp_size=2)`.
Six tests say it that way now. This is the price of having one answer computed
once, and it is the same price every other derived value in the config already
carries.

A caller that reads a quotient without publishing or overriding now gets an
explicit error naming the field, instead of a default that an uninitialised
group happened to supply. One fixture was in that state --
`TestMlaWriteDoorsUnderDcp` built a bare pool and asked whether DCP was on --
and it publishes a config now, which is what the process it stands in for
does.

Eighteen sites read these predicates without calling them. That is correct --
they are properties -- but it is worth saying they were checked, because a
census that assumes otherwise reports eighteen always-true conditions.

## The skill that documents this subsystem is updated with it

`.claude/rules/modify-component-must-read.md` points at
`.claude/skills/sglang-runtime-context/SKILL.md` before anyone touches these
files, so a stale sentence there is a wrong instruction rather than a stale
note. Four of its load-bearing statements stopped being true across this series
and are corrected here: `NS(...)` is no longer how a field states its namespace
(the declaring class is); the DCP degrade rule is gone, because the quotients
are not live reads; `mamba_extra_buffer_enabled()` and the other predicate
functions it named as the shape to copy no longer exist; and the
namespace-coverage ratchet is described in terms of the marker. The docstring of
`test_server_args_namespaces.py` said the same thing and is fixed too.

The consequence a test author actually trips over is stated there as well:
overriding a leaf no longer moves its quotient, so a topology is stated by
publishing a config or by naming the width.

## Verification

A full registered-unit sweep (648 files) against this stack's merge-base:
19 failures on both sides, the same 19 -- AMD `gfx950`, `modelopt`,
`cuda_vmm`, `weight_checker` and friends, none of them config. The narrower 139-file config sweep used earlier in this series
does not contain the files this change reaches -- `test_kv_index_translator`
never names `get_parallel()`, it constructs an object that does -- which is why
the baseline differential over everything is what is quoted here.
2026-09-06 21:44:24 -07:00
sglang-botandsglang-bot 6252993afe chore: update CI test est_time values (#38238)
Co-authored-by: sglang-bot <sglang-bot@users.noreply.github.com>
2026-09-06 17:49:41 -07:00
cctryandcctry 9a05b470fa [Memory] Size the CUDA graph pool from warmup measurements and fix graph-pool borrowing (#36911)
Co-authored-by: cctry <cctry@fb.com>
2026-09-01 09:32:38 -07:00
YAMY 33ed29a0ee test: update hybrid attention runner fixtures (#37345) 2026-08-31 22:00:50 -07:00
YAMYandgithub-actions[bot] b77cac06a9 [PP] Support prefill CUDA graph proxy tensors (#36248)
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
2026-08-30 23:11:45 -07:00
+8 5f216fc33f qwen 3.8 rebase (#35758)
Co-authored-by: cherichy <cherichy@outlook.com>
Co-authored-by: guangyunh-nv <guangyunh@nvidia.com>
Co-authored-by: jiahanc <jiahanc@nvidia.com>
Co-authored-by: jinyangyuan-nvidia <joyuan@nvidia.com>
Co-authored-by: Cheng Hang <chang@nvidia.com>
Co-authored-by: Yicheng Qiang <yqiang@nvidia.com>
Co-authored-by: Sam Li <lsam@nvidia.com>
Co-authored-by: Tom-Zheng <tizheng@nvidia.com>
Co-authored-by: Yangmin Li <yangminl@nvidia.com>
Co-authored-by: xiaoweiw-nv <xiaoweiw@nvidia.com>
Co-authored-by: Zheng Li <lizheng.cs@zju.edu.cn>
Co-authored-by: yizhang2077 <1109276519@qq.com>
Co-authored-by: Ke Bao <ispobaoke@gmail.com>
Co-authored-by: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com>
Co-authored-by: Yuhao Yang <47235274+yhyang201@users.noreply.github.com>
Co-authored-by: Zijie Xia <zijie.xia@radixark.ai>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
2026-08-28 20:41:34 -07:00
Aurick QiaoandYuwei An 26fd7fdaa2 [Feature] Support PP in full prefill CUDA graphs (#35451)
Co-authored-by: Yuwei An <ayw.sirius19@gmail.com>
2026-08-27 17:32:00 -07:00
Cheng Wan ae5feb4b9c config: stop handing the record to code that does not read it (#36252) 2026-08-26 05:02:17 -07:00
092d85eb87 [Feature] Add MiniCPM-SALA support (#30360)
Co-authored-by: Alex Nails <alex.nails@radixark.ai>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-24 02:25:16 -07:00
Cheng WanandClaude Opus 5 362c2ee849 config: borrowed-record reads follow the config bags (#35908)
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
2026-08-23 01:19:20 -07:00
c14312a664 [Spec] DFlash2: local convolution + candidate selector (#35371)
Co-authored-by: Jian Chen <jianchen0311@gmail.com>
Co-authored-by: Liangsheng Yin <hnyls2002@gmail.com>
Co-authored-by: hnyls2002 <lsyincs@gmail.com>
2026-08-18 17:07:28 -07:00
Cheng Wan a97bc8db32 config: the DP/EP topology reads come from the parallel bag (#35025) 2026-08-17 16:17:26 -07:00
Mohammad Miadh Angkad d7207be156 Fix startup weight load after TorchAO removal (#34869) 2026-08-14 13:18:53 -07:00
6b94d39f13 [Model Loading] Overlap checkpoint staging with CUDA graph capture during startup (#32017)
Co-authored-by: Wenhui Zhu <wzhu59@asu.edu>
Co-authored-by: Alex Nails <alex.nails@radixark.ai>
2026-08-13 12:26:25 -07:00
Mick db75dfe10f fix: always capture default prefill CUDA graph (#33352) 2026-08-08 19:24:49 +08:00
Lianmin ZhengandItai Gat dea2be5ae3 [CUDA Graph] Allow custom decode graph runners (#33553)
Co-authored-by: Itai Gat <itaigat.mail@gmail.com>
2026-08-04 12:48:56 -07:00
Peng Wu e3d4f48e55 [Fix] missing max_context_len on HybridAttnBackend (#32690) 2026-07-31 19:43:09 +08:00
Mohammad Miadh Angkad 0bdd4730af [CI] Fix failures on main (#32091) 2026-07-23 00:11:25 +08:00
ThanhhaoandHao Phan 72c4ed1a3f [Spec] DFlash: remove per-step host syncs so the CPU runs a full step ahead (spec-v2 overlap) (#31468)
Co-authored-by: Hao Phan <htphan@nvidia.com>
2026-07-17 23:22:07 -07:00
Mick 681c223570 refactor: wrap split backends once on full-attention backends (#31439) 2026-07-17 19:15:04 +08:00
Mick 24a8944e15 fix: enable Kimi multimodal breakable prefill cuda graph replay (#31391) 2026-07-17 19:13:54 +08:00
Mick d9003dd452 fix: skip unsafe automatic prefill graph capture (#31204) 2026-07-16 09:27:38 +08:00
fzyzcjy d15f6a9ac3 Introduce NgramEmbeddingManager component (#31154) 2026-07-14 15:58:08 +08:00