Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions .github/workflows/test.yml
Original file line number Diff line number Diff line change
Expand Up @@ -91,6 +91,8 @@ jobs:
if loaded:
sys.exit(f"public SDK import loaded private implementation: {loaded}")
caps = get_capabilities()
if "vllm-gr" not in caps.supported_repositories:
sys.exit("installed SDK omits vLLM-GR review support")
dist_version = importlib.metadata.version("infermatrix-copilot")
if not (__version__ == dist_version == caps.distribution_version == "0.2.0"):
sys.exit(
Expand Down
118 changes: 118 additions & 0 deletions adapters/vllm_gr/manifest.yaml
Original file line number Diff line number Diff line change
@@ -0,0 +1,118 @@
# GR review support authorized by the repository maintainer.
name: vllm_gr
status: active
created_by: human
schema_version: 1
capabilities: []
repo:
path: ${VLLM_GR_REPO}
full_name: JiusiServe/vllm-gr
aliases:
- vllm-gr
- vllm_gr
- JiusiServe/vllm-gr

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Badge Allow the declared full-name alias in the SDK

When an SDK caller uses the newly declared JiusiServe/vllm-gr alias as RepositoryRef.alias, DirectClient._request_values() rejects it as unsupported before adapter normalization because it compares only against canonical names from supported_repositories. The MCP path accepts this identifier, and the parity test masks the discrepancy by passing vllm-gr to the SDK but the full name only to direct_review_plan(); either make the SDK honor adapter aliases or stop advertising this value as an alias.

Useful? React with 👍 / 👎.

default_branch: main
language: python
knowledge:
source: JiusiServe/vllm-gr@a05103dc16c7b8dfd2213b73c5ee2d4f28cb6fee
repo_subdir: repos/vllm-gr
briefing_docs:
- repos/vllm-gr/_index.md
modules:
engine_core:
local_paths:
- vllm_gr/v1/engine/
- vllm_gr/v1/core/
wave: 1
risk: high
beam_proc:
local_paths:
- rust/beam-proc/
- rust/ce-transport/
- rust/examples/
wave: 1
risk: high
serving:
local_paths:
- vllm_gr/entrypoints/
wave: 1
risk: high
attention:
local_paths:
- vllm_gr/v1/attention/
wave: 1
risk: high
packaging:
local_paths:
- setup.py
- pyproject.toml
- MANIFEST.in
- requirements/
- rust/Cargo.
- .github/workflows/
- .pre-commit-config.yaml
- tools/pre_commit/
- tests/packaging/
wave: 1
risk: high
review_routes:
- prefix: vllm_gr/v1/engine/
owner: engine-core
doc: repos/vllm-gr/components/engine-core/architecture.md
- prefix: vllm_gr/v1/core/
owner: engine-core
doc: repos/vllm-gr/components/engine-core/architecture.md
- prefix: rust/beam-proc/
owner: beam-proc
doc: repos/vllm-gr/components/beam-proc/architecture.md
- prefix: rust/ce-transport/
owner: beam-proc
doc: repos/vllm-gr/components/beam-proc/architecture.md
- prefix: rust/examples/
owner: beam-proc
doc: repos/vllm-gr/components/beam-proc/architecture.md
- prefix: vllm_gr/entrypoints/
owner: serving
doc: repos/vllm-gr/components/serving/architecture.md
- prefix: vllm_gr/v1/attention/
owner: attention
doc: repos/vllm-gr/components/attention/architecture.md
- prefix: setup.py
owner: packaging
doc: repos/vllm-gr/components/packaging/architecture.md
- prefix: pyproject.toml
owner: packaging
doc: repos/vllm-gr/components/packaging/architecture.md
- prefix: MANIFEST.in
owner: packaging
doc: repos/vllm-gr/components/packaging/architecture.md
- prefix: requirements/
owner: packaging
doc: repos/vllm-gr/components/packaging/architecture.md
- prefix: rust/Cargo.
owner: packaging
doc: repos/vllm-gr/components/packaging/architecture.md
- prefix: .github/workflows/
owner: packaging
doc: repos/vllm-gr/components/packaging/architecture.md
- prefix: .pre-commit-config.yaml
owner: packaging
doc: repos/vllm-gr/components/packaging/architecture.md
- prefix: tools/pre_commit/
owner: packaging
doc: repos/vllm-gr/components/packaging/architecture.md
- prefix: tests/packaging/
owner: packaging
doc: repos/vllm-gr/components/packaging/architecture.md
validation:
test_manifest_sources:
- .github/workflows/
precommit: true
ci:
provider: github_actions
push:
default_remote: origin
protected_branches:
- main
- master
allowed: false
2 changes: 1 addition & 1 deletion doc/knowledge/SCHEMA.md
Original file line number Diff line number Diff line change
Expand Up @@ -45,7 +45,7 @@ contradictions: [相对路径] # 可选:与本页冲突的页面

新标签必须先加入此表再使用(防止标签蔓延):

- 归属:`general`、`vllm-omni`、`afd-plugin`
- 归属:`general`、`vllm-omni`、`afd-plugin`、`vllm-gr`
- 工作主题:`review`、`ci`、`docs`、`git`、`debug`、`bug`、`benchmark`、`environment`、
`remote`、`agents`、`planning`、`dev`、`rebase`
- 代码/模型轴:`components`、`models`、`diffusion`、`model-executor`、`serving`、
Expand Down
5 changes: 3 additions & 2 deletions knowledge/repos/_index.md
Original file line number Diff line number Diff line change
@@ -1,9 +1,9 @@
---
title: "仓库经验入口"
created: 2026-07-10
updated: 2026-08-06
updated: 2026-09-30
type: index
tags: [vllm-omni, afd-plugin]
tags: [vllm-omni, afd-plugin, vllm-gr]
sources: []
---

Expand All @@ -26,3 +26,4 @@ sources: []
| vLLM-Omni | `vllm-project/vllm-omni` | [vllm-omni](vllm-omni/_index.md) |
| afd-plugin | `vllm-project/afd-plugin` | [afd-plugin](afd-plugin/_index.md) |
| Jianghan roleplay data pipeline | 私有项目 | [jianghan-roleplay-data-pipeline](jianghan-roleplay-data-pipeline/_index.md) |
| vLLM-GR | `JiusiServe/vllm-gr` | [vllm-gr](vllm-gr/_index.md) |
33 changes: 33 additions & 0 deletions knowledge/repos/vllm-gr/_index.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,33 @@
---
title: vLLM-GR repository entry
created: '2026-09-30'
updated: '2026-09-30'
type: index
tags:
- vllm-gr
sources:
- JiusiServe/vllm-gr@a05103dc16c7b8dfd2213b73c5ee2d4f28cb6fee:README.md
---

# vLLM-GR repository entry

## When to use this entry

Repository: `JiusiServe/vllm-gr`; aliases `vllm-gr` and `vllm_gr`.
Default branch: `main`. Source map baseline: `a05103dc16c7b8dfd2213b73c5ee2d4f28cb6fee`.
The baseline supports vLLM `0.22.1` and vLLM-Ascend `0.22.1rc1`; inspect the target revision's own constraints before validation.
This repository implements generative recommendation inference, including OneRec beam/catalog paths, Pangu passthrough and GPU/NPU execution.

## Outside this entry

Repository-independent review procedures belong to [general review](../../general/review/_index.md).
vLLM-Omni model and diffusion assumptions belong to that repository and do not apply here.
This map describes the pinned main branch; new transports and APIs must be verified from the reviewed head.

## Contents

| Task | Entry | Boundary |
|---|---|---|
| Select a source owner and focused tests | [Components](components/_index.md) | EngineCore, BeamProc, serving, attention and packaging |

Direct owner signals and source scopes are declared in `_routes.yaml`; adapter aliases and briefing live in `adapters/vllm_gr/manifest.yaml`.
75 changes: 75 additions & 0 deletions knowledge/repos/vllm-gr/_routes.yaml
Original file line number Diff line number Diff line change
@@ -0,0 +1,75 @@
schema_version: 1
owners:
- owner: engine-core
path: repos/vllm-gr/components/engine-core/architecture.md
signals:
- enginecore
- engine core
- scheduler
- scheduling
- beam plan
- beam cache
- ce
scope_prefixes:
- vllm_gr/v1/engine/
- vllm_gr/v1/core/
- owner: beam-proc
path: repos/vllm-gr/components/beam-proc/architecture.md
signals:
- beamproc
- beam proc
- bp
- rust worker
- shared memory
- sharedmem
- iceoryx
- transport
scope_prefixes:
- rust/beam-proc/
- rust/ce-transport/
- rust/examples/
- owner: serving
path: repos/vllm-gr/components/serving/architecture.md
signals:
- serving
- openai
- chat completion
- streaming
- sse
- beam stream
- beam metrics
- response
scope_prefixes:
- vllm_gr/entrypoints/
- owner: attention
path: repos/vllm-gr/components/attention/architecture.md
signals:
- attention
- flash attention
- flashattn
- suffix lse
- cascade attention
- gqa
- kv cache
scope_prefixes:
- vllm_gr/v1/attention/
Comment on lines +54 to +55

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Badge Add focused test paths to owner scopes

When a PR has a generic title and changes only the documented focused test—for example tests/test_beam_attention_metadata.py—none of these scope prefixes match, so direct_review_plan() returns no knowledge route (needs_pr_context) instead of the attention owner. The same gap affects the focused engine, BeamProc, and serving tests, while packaging already routes tests/packaging/; add the stable focused test paths to their corresponding owner scopes so test-only GR reviews receive the relevant review map.

Useful? React with 👍 / 👎.

- owner: packaging
path: repos/vllm-gr/components/packaging/architecture.md
signals:
- packaging
- wheel
- cargo
- native library
- build
- pre commit
- installation
scope_prefixes:
- setup.py
- pyproject.toml
- MANIFEST.in
- requirements/
- rust/Cargo.
- .github/workflows/
- .pre-commit-config.yaml
- tools/pre_commit/
- tests/packaging/
30 changes: 30 additions & 0 deletions knowledge/repos/vllm-gr/components/_index.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,30 @@
---
title: vLLM-GR component map
created: '2026-09-30'
updated: '2026-09-30'
type: index
tags:
- vllm-gr
- components
sources: []
---

# vLLM-GR components

## When to use this entry

Select the stable source owner for a GR runtime or packaging change.

## Outside this map

Repository identity and supported versions belong to the [repository entry](../_index.md); generic procedures belong to [general review](../../../general/review/_index.md).

## Contents

| Task | Entry | Source scope |
|---|---|---|
| EngineCore beam scheduling | [engine-core](engine-core/_index.md) | `vllm_gr/v1/engine/` |
| Rust BeamProc runtime | [beam-proc](beam-proc/_index.md) | `rust/beam-proc/` |
| OpenAI beam serving | [serving](serving/_index.md) | `vllm_gr/entrypoints/` |
| Beam attention runtime | [attention](attention/_index.md) | `vllm_gr/v1/attention/` |
| Native packaging and validation | [packaging](packaging/_index.md) | `setup.py` |
32 changes: 32 additions & 0 deletions knowledge/repos/vllm-gr/components/attention/_index.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,32 @@
---
title: Beam attention runtime entry
created: '2026-09-30'
updated: '2026-09-30'
type: index
tags:
- vllm-gr
- attention-runtime
sources:
- JiusiServe/vllm-gr@a05103dc16c7b8dfd2213b73c5ee2d4f28cb6fee:vllm_gr/v1/attention/backends/beam_attn_gpu.py
- JiusiServe/vllm-gr@a05103dc16c7b8dfd2213b73c5ee2d4f28cb6fee:vllm_gr/v1/attention/backends/beam_attn_metadata.py
---

# Beam attention runtime

## When to use this entry

GPU/NPU beam attention execution and the metadata that describes ordinary, prefix and suffix query/cache layouts.

Source scopes: `vllm_gr/v1/attention/`.

## Outside this owner

Other source owners are listed in the [component map](../_index.md); repository identity and version constraints belong to the [repository entry](../../_index.md).

## Contents

| Task | Entry | Purpose |
|---|---|---|
| Trace source, data flow and validation | [Architecture](architecture.md) | Pinned source map and compact Direct review map |

Focused tests: `tests/test_beam_attention_metadata.py`.
39 changes: 39 additions & 0 deletions knowledge/repos/vllm-gr/components/attention/architecture.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,39 @@
---
title: Beam attention runtime architecture
created: '2026-09-30'
updated: '2026-09-30'
type: architecture
tags:
- vllm-gr
- attention-runtime
sources:
- JiusiServe/vllm-gr@a05103dc16c7b8dfd2213b73c5ee2d4f28cb6fee:vllm_gr/v1/attention/backends/beam_attn_gpu.py
- JiusiServe/vllm-gr@a05103dc16c7b8dfd2213b73c5ee2d4f28cb6fee:vllm_gr/v1/attention/backends/beam_attn_metadata.py
---

# Beam attention runtime architecture

## Responsibilities and boundary

GPU/NPU beam attention execution and the metadata that describes ordinary, prefix and suffix query/cache layouts.

## Source and entrypoints

- `vllm_gr/v1/attention/backends/beam_attn_gpu.py`
- `vllm_gr/v1/attention/backends/beam_attn_metadata.py`

## Data flow

Metadata partitions standard and beam queries and supplies cache geometry. GPU attention evaluates standard queries and prefix/suffix beam attention, merges their states and writes into the caller-owned output. Platform and FlashAttention-version branches can have different shape and ownership contracts.

## Direct review map

Read `vllm_gr/v1/attention/backends/beam_attn_metadata.py` for query indices and cache geometry, then `beam_attn_gpu.py` for prefix/suffix kernels, LSE layouts and output writes. Check ordinary and grouped queries, supported FlashAttention versions, causal suffix lengths and caller output ownership. Inspect platform-specific peers before extending a GPU conclusion to NPU.

Focused tests: `tests/test_beam_attention_metadata.py`.

## Validation

Use the target head's declared dependency versions and test selectors. Bind evidence to that head; a missing native build or incompatible environment is a validation gap.

See the [owner entry](_index.md) and [component map](../_index.md) for neighboring boundaries.
Loading
Loading