repo stringlengths 5 53 | pr_number int32 1 321k | task_type stringclasses 2
values | issue_text stringlengths 0 81.2k | pr_title stringlengths 1 319 | pr_body stringlengths 0 105k | base_sha stringlengths 40 40 | head_sha stringlengths 40 40 | gold_diff stringlengths 0 202M | changed_files listlengths 0 100 | review_threads listlengths 0 100 | test_patch stringlengths 0 23.4M | merged bool 1
class |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
vitejs/vite | 21,906 | issue_to_patch | feat: add warning to discourage Vite with yarn pnp | Adding the following message to discourage people from using yarn pnp with Vite:
> Using Yarn PnP with Vite is discouraged and PnP-specific bugs will no longer be actively worked on.
> Please switch to a different **'nodeLinker'** mode or to a different package manager.
While yarn defaults to yarn pnp when runn... | dc245c71e5007ea4d891a025e2d69ac96c736546 | 833b40798675fd0d9eb9d15cac7826680d0cdbb6 | diff --git a/packages/vite/src/node/config.ts b/packages/vite/src/node/config.ts
index 211217defd62ef..e975d718d53ad5 100644
--- a/packages/vite/src/node/config.ts
+++ b/packages/vite/src/node/config.ts
@@ -1480,6 +1480,15 @@ export async function resolveConfig(
)
}
+ if (process.versions.pnp) {
+ logger.... | [
"packages/vite/src/node/config.ts"
] | [] | true | ||
vitejs/vite | 21,412 | issue_to_patch | Asset discovery for custom HTML elements and/or attributes
### Description
As a developer who makes extensive use of web components and ad hoc HTML attributes, I would like to see Vite support asset resolution/discovery for arbitrary elements and attributes. Some of my use cases:
```html
<!-- Web component refering ... | feat(html): add `html.additionalAssetSources` option | ## Issue
close #21182
## Description
This PR adds the `html.additionalAssetSources` option to allow users to configure custom HTML elements and attributes as asset sources.
Currently, Vite only processes assets in hardcoded HTML elements and attributes (`<img src>`, `<video src>`, etc.). Users cannot configure... | dc245c71e5007ea4d891a025e2d69ac96c736546 | 743adad8fea9c0ebf4d530a1e9c3069cb0352a23 | diff --git a/docs/config/shared-options.md b/docs/config/shared-options.md
index da4cb198adc704..dd6a4d126f2bc8 100644
--- a/docs/config/shared-options.md
+++ b/docs/config/shared-options.md
@@ -214,6 +214,48 @@ Enables the tsconfig paths resolution feature. `paths` option in `tsconfig.json`
A nonce value placeholde... | [
"docs/config/shared-options.md",
"packages/vite/src/node/__tests__/assetSource.spec.ts",
"packages/vite/src/node/assetSource.ts",
"packages/vite/src/node/config.ts",
"packages/vite/src/node/index.ts",
"packages/vite/src/node/plugins/html.ts",
"packages/vite/src/node/server/middlewares/indexHtml.ts"
] | [
{
"comment": "If the config is:\n\n```js\n'img': { srcAttributes: ['data-src-dark', 'data-src-light'] }\n```\n\nDoesn't this code only supports `data-` and no longer `src`? We probably need to run the matching twice here if `additionalAssetSources` is defined.",
"path": "packages/vite/src/node/assetSource.t... | diff --git a/packages/vite/src/node/__tests__/assetSource.spec.ts b/packages/vite/src/node/__tests__/assetSource.spec.ts
index 2c05d90e57b265..66a857b92886ff 100644
--- a/packages/vite/src/node/__tests__/assetSource.spec.ts
+++ b/packages/vite/src/node/__tests__/assetSource.spec.ts
@@ -94,4 +94,103 @@ describe('getNode... | true |
vitejs/vite | 21,412 | comment_to_fix | feat(html): add `html.additionalAssetSources` option | If the config is:
```js
'img': { srcAttributes: ['data-src-dark', 'data-src-light'] }
```
Doesn't this code only supports `data-` and no longer `src`? We probably need to run the matching twice here if `additionalAssetSources` is defined. | dc245c71e5007ea4d891a025e2d69ac96c736546 | 743adad8fea9c0ebf4d530a1e9c3069cb0352a23 | diff --git a/packages/vite/src/node/assetSource.ts b/packages/vite/src/node/assetSource.ts
index b733258896678f..d69027e5ed48c6 100644
--- a/packages/vite/src/node/assetSource.ts
+++ b/packages/vite/src/node/assetSource.ts
@@ -1,7 +1,19 @@
import type { DefaultTreeAdapterMap, Token } from 'parse5'
-interface HtmlAss... | [
"packages/vite/src/node/assetSource.ts"
] | [
{
"comment": "If the config is:\n\n```js\n'img': { srcAttributes: ['data-src-dark', 'data-src-light'] }\n```\n\nDoesn't this code only supports `data-` and no longer `src`? We probably need to run the matching twice here if `additionalAssetSources` is defined.",
"path": "packages/vite/src/node/assetSource.t... | true | ||
vitejs/vite | 21,412 | comment_to_fix | feat(html): add `html.additionalAssetSources` option | I think we should run it twice instead. The additional `filter` would apply only for the object specified in `additionalAssetSources`, so we can't merge it. Plus merging/spreading is a bit expensive if the html have many references. | dc245c71e5007ea4d891a025e2d69ac96c736546 | 743adad8fea9c0ebf4d530a1e9c3069cb0352a23 | diff --git a/packages/vite/src/node/assetSource.ts b/packages/vite/src/node/assetSource.ts
index b733258896678f..d69027e5ed48c6 100644
--- a/packages/vite/src/node/assetSource.ts
+++ b/packages/vite/src/node/assetSource.ts
@@ -1,7 +1,19 @@
import type { DefaultTreeAdapterMap, Token } from 'parse5'
-interface HtmlAss... | [
"packages/vite/src/node/assetSource.ts"
] | [
{
"comment": "I think we should run it twice instead. The additional `filter` would apply only for the object specified in `additionalAssetSources`, so we can't merge it. Plus merging/spreading is a bit expensive if the html have many references.",
"path": "packages/vite/src/node/assetSource.ts",
"hunk"... | true | ||
vitejs/vite | 21,205 | issue_to_patch | Refactor to reference `rolldownOptions` mainly instead of `rollupOptions`
Currently the code base references `rollupOptions` instead of `rolldownOptions`, but we should be using `rolldownOptions` in most places.
fix: refer to `rolldownOptions.input` instead of deprecated `rollupOptions.input` in messages (missed by #... | refactor: use rolldownOptions where possible | ### Description
Replaces all legacy `rollupOptions` calls with `rolldownOptions` (except compat layer)
| dc245c71e5007ea4d891a025e2d69ac96c736546 | ee257c324881a8d96b768b83ceff1ec1768f11b9 | diff --git a/docs/guide/backend-integration.md b/docs/guide/backend-integration.md
index 7cce7dbcde1492..6699a8d157d3d2 100644
--- a/docs/guide/backend-integration.md
+++ b/docs/guide/backend-integration.md
@@ -179,7 +179,7 @@ If you need a custom integration, you can follow the steps in this guide to conf
```
... | [
"docs/guide/backend-integration.md",
"packages/plugin-legacy/src/index.ts",
"packages/vite/src/node/__tests__/build.spec.ts",
"packages/vite/src/node/__tests__/config.spec.ts",
"packages/vite/src/node/__tests__/dev.spec.ts",
"packages/vite/src/node/__tests__/environment.spec.ts",
"packages/vite/src/node... | [
{
"comment": "This was removed in another PR and shouldn't be needed. Probably a merge issue.",
"path": "packages/vite/src/node/build.ts",
"hunk": "@@ -620,39 +620,49 @@ export function resolveRolldownOptions(\n injectEnvironmentToHooks(environment, chunkMetadataMap, p),\n )\n \n- const rollupOpt... | diff --git a/packages/vite/src/node/__tests__/build.spec.ts b/packages/vite/src/node/__tests__/build.spec.ts
index 141fc682f9edbd..b8dcb7eb035256 100644
--- a/packages/vite/src/node/__tests__/build.spec.ts
+++ b/packages/vite/src/node/__tests__/build.spec.ts
@@ -153,7 +153,7 @@ describe('build', () => {
write:... | true |
vitejs/vite | 21,205 | comment_to_fix | refactor: use rolldownOptions where possible | This was removed in another PR and shouldn't be needed. Probably a merge issue. | dc245c71e5007ea4d891a025e2d69ac96c736546 | ee257c324881a8d96b768b83ceff1ec1768f11b9 | diff --git a/packages/vite/src/node/build.ts b/packages/vite/src/node/build.ts
index 870c0eb1bafdec..133cbc3d2da5ce 100644
--- a/packages/vite/src/node/build.ts
+++ b/packages/vite/src/node/build.ts
@@ -195,7 +195,7 @@ export interface BuildEnvironmentOptions {
rollupOptions?: RolldownOptions
/**
* Will be me... | [
"packages/vite/src/node/build.ts"
] | [
{
"comment": "This was removed in another PR and shouldn't be needed. Probably a merge issue.",
"path": "packages/vite/src/node/build.ts",
"hunk": "@@ -620,39 +620,49 @@ export function resolveRolldownOptions(\n injectEnvironmentToHooks(environment, chunkMetadataMap, p),\n )\n \n- const rollupOpt... | true | ||
vitejs/vite | 21,205 | comment_to_fix | refactor: use rolldownOptions where possible | I think the previous one is correct | dc245c71e5007ea4d891a025e2d69ac96c736546 | ee257c324881a8d96b768b83ceff1ec1768f11b9 | diff --git a/packages/vite/src/node/config.ts b/packages/vite/src/node/config.ts
index b3a29711cd45cd..411a1b0dda5a8c 100644
--- a/packages/vite/src/node/config.ts
+++ b/packages/vite/src/node/config.ts
@@ -858,7 +858,7 @@ const configDefaults = Object.freeze({
worker: {
format: 'iife',
plugins: (): never[... | [
"packages/vite/src/node/config.ts"
] | [
{
"comment": "I think the previous one is correct",
"path": "packages/vite/src/node/config.ts",
"hunk": "@@ -1829,8 +1829,8 @@ export async function resolveConfig(\n } = {\n format: config.worker?.format || 'iife',\n plugins: createWorkerPlugins,\n- rollupOptions: config.worker?.rollupOptio... | true | ||
vitejs/vite | 21,205 | comment_to_fix | refactor: use rolldownOptions where possible | ```suggestion
...rolldownOptions.transform,
``` | dc245c71e5007ea4d891a025e2d69ac96c736546 | ee257c324881a8d96b768b83ceff1ec1768f11b9 | diff --git a/packages/vite/src/node/plugins/worker.ts b/packages/vite/src/node/plugins/worker.ts
index 25eedfc7b1aac5..240a2ebba0b8cd 100644
--- a/packages/vite/src/node/plugins/worker.ts
+++ b/packages/vite/src/node/plugins/worker.ts
@@ -183,7 +183,7 @@ async function bundleWorkerEntry(
// bundle the file as entr... | [
"packages/vite/src/node/plugins/worker.ts"
] | [
{
"comment": "```suggestion\n ...rolldownOptions.transform,\n```",
"path": "packages/vite/src/node/plugins/worker.ts",
"hunk": "",
"resolving_sha": "ee257c324881a8d96b768b83ceff1ec1768f11b9",
"resolving_diff": "diff --git a/packages/vite/src/node/plugins/worker.ts b/packages/vite/src/node/p... | true | ||
vitejs/vite | 21,205 | comment_to_fix | refactor: use rolldownOptions where possible | ```suggestion
...rolldownOptions.transform?.define,
``` | dc245c71e5007ea4d891a025e2d69ac96c736546 | ee257c324881a8d96b768b83ceff1ec1768f11b9 | diff --git a/packages/vite/src/node/plugins/worker.ts b/packages/vite/src/node/plugins/worker.ts
index 25eedfc7b1aac5..240a2ebba0b8cd 100644
--- a/packages/vite/src/node/plugins/worker.ts
+++ b/packages/vite/src/node/plugins/worker.ts
@@ -183,7 +183,7 @@ async function bundleWorkerEntry(
// bundle the file as entr... | [
"packages/vite/src/node/plugins/worker.ts"
] | [
{
"comment": "```suggestion\n ...rolldownOptions.transform?.define,\n```",
"path": "packages/vite/src/node/plugins/worker.ts",
"hunk": "",
"resolving_sha": "ee257c324881a8d96b768b83ceff1ec1768f11b9",
"resolving_diff": "diff --git a/packages/vite/src/node/plugins/worker.ts b/packages/vite/... | true | ||
vitejs/vite | 22,584 | issue_to_patch | ci: rewrite issue-helper with JS script | Rewrote issue-helper with JS script as it's been disabled for a while.
Based on https://github.com/vitest-dev/vitest/pull/10416, https://github.com/vitest-dev/vitest/pull/10455, https://github.com/vitejs/vite-plugin-react/pull/1238
| f94df87ff03b40b65e29bacdc04cc18c7bccaa4a | de76b899686df250779de2bfe88388750e9a84ea | diff --git a/.github/actions/issues-helper/action.yml b/.github/actions/issues-helper/action.yml
new file mode 100644
index 00000000000000..f2aeb0937f2cc9
--- /dev/null
+++ b/.github/actions/issues-helper/action.yml
@@ -0,0 +1,168 @@
+name: Issues Helper
+description: Add labels, create comments, close stale issues, or... | [
".github/actions/issues-helper/action.yml",
".github/workflows/issue-close-require.yml",
".github/workflows/issue-labeled.yml",
".github/workflows/lock-closed-issues.yml"
] | [
{
"comment": "`inputs.actions` documentation is missing `remove-labels`, even though the composite action implements it (and workflows use it). This makes the action interface misleading for future callers.",
"path": ".github/actions/issues-helper/action.yml",
"hunk": "@@ -0,0 +1,168 @@\n+name: Issues H... | true | ||
vitejs/vite | 22,584 | comment_to_fix | ci: rewrite issue-helper with JS script | `inputs.actions` documentation is missing `remove-labels`, even though the composite action implements it (and workflows use it). This makes the action interface misleading for future callers. | f94df87ff03b40b65e29bacdc04cc18c7bccaa4a | de76b899686df250779de2bfe88388750e9a84ea | diff --git a/.github/actions/issues-helper/action.yml b/.github/actions/issues-helper/action.yml
new file mode 100644
index 00000000000000..f2aeb0937f2cc9
--- /dev/null
+++ b/.github/actions/issues-helper/action.yml
@@ -0,0 +1,168 @@
+name: Issues Helper
+description: Add labels, create comments, close stale issues, or... | [
".github/actions/issues-helper/action.yml"
] | [
{
"comment": "`inputs.actions` documentation is missing `remove-labels`, even though the composite action implements it (and workflows use it). This makes the action interface misleading for future callers.",
"path": ".github/actions/issues-helper/action.yml",
"hunk": "@@ -0,0 +1,168 @@\n+name: Issues H... | true | ||
vitejs/vite | 22,584 | comment_to_fix | ci: rewrite issue-helper with JS script | `issue-number` is also required for `remove-labels` (the script calls `issues.removeLabel` for the provided issue/PR). The current description only mentions `add-labels`/`create-comment`, which can lead to misconfiguration. | f94df87ff03b40b65e29bacdc04cc18c7bccaa4a | de76b899686df250779de2bfe88388750e9a84ea | diff --git a/.github/actions/issues-helper/action.yml b/.github/actions/issues-helper/action.yml
new file mode 100644
index 00000000000000..f2aeb0937f2cc9
--- /dev/null
+++ b/.github/actions/issues-helper/action.yml
@@ -0,0 +1,168 @@
+name: Issues Helper
+description: Add labels, create comments, close stale issues, or... | [
".github/actions/issues-helper/action.yml"
] | [
{
"comment": "`issue-number` is also required for `remove-labels` (the script calls `issues.removeLabel` for the provided issue/PR). The current description only mentions `add-labels`/`create-comment`, which can lead to misconfiguration.",
"path": ".github/actions/issues-helper/action.yml",
"hunk": "@@ ... | true | ||
vitejs/vite | 22,584 | comment_to_fix | ci: rewrite issue-helper with JS script | `labels` is also required for `remove-labels` (and is consumed by that step). The current description omits that, which makes the action inputs harder to use correctly. | f94df87ff03b40b65e29bacdc04cc18c7bccaa4a | de76b899686df250779de2bfe88388750e9a84ea | diff --git a/.github/actions/issues-helper/action.yml b/.github/actions/issues-helper/action.yml
new file mode 100644
index 00000000000000..f2aeb0937f2cc9
--- /dev/null
+++ b/.github/actions/issues-helper/action.yml
@@ -0,0 +1,168 @@
+name: Issues Helper
+description: Add labels, create comments, close stale issues, or... | [
".github/actions/issues-helper/action.yml"
] | [
{
"comment": "`labels` is also required for `remove-labels` (and is consumed by that step). The current description omits that, which makes the action inputs harder to use correctly.",
"path": ".github/actions/issues-helper/action.yml",
"hunk": "@@ -0,0 +1,168 @@\n+name: Issues Helper\n+description: Add... | true | ||
vitejs/vite | 22,584 | comment_to_fix | ci: rewrite issue-helper with JS script | The `close-issues` action can unintentionally close pull requests (since `issues.listForRepo` returns PRs too) and can close issues that have recent activity, because it only checks when the label was applied. This conflicts with the intended "no activity within N days" behavior and risks closing active threads. | f94df87ff03b40b65e29bacdc04cc18c7bccaa4a | de76b899686df250779de2bfe88388750e9a84ea | diff --git a/.github/actions/issues-helper/action.yml b/.github/actions/issues-helper/action.yml
new file mode 100644
index 00000000000000..f2aeb0937f2cc9
--- /dev/null
+++ b/.github/actions/issues-helper/action.yml
@@ -0,0 +1,168 @@
+name: Issues Helper
+description: Add labels, create comments, close stale issues, or... | [
".github/actions/issues-helper/action.yml"
] | [
{
"comment": "The `close-issues` action can unintentionally close pull requests (since `issues.listForRepo` returns PRs too) and can close issues that have recent activity, because it only checks when the label was applied. This conflicts with the intended \"no activity within N days\" behavior and risks closin... | true | ||
vitejs/vite | 22,584 | comment_to_fix | ci: rewrite issue-helper with JS script | This workflow only runs on `issues` events (not PRs), but it still requests `pull-requests: write`. Dropping unnecessary token permissions reduces blast radius if the token is ever misused. | f94df87ff03b40b65e29bacdc04cc18c7bccaa4a | de76b899686df250779de2bfe88388750e9a84ea | diff --git a/.github/workflows/issue-labeled.yml b/.github/workflows/issue-labeled.yml
index 54536cc342387a..a52e5cc527e06a 100644
--- a/.github/workflows/issue-labeled.yml
+++ b/.github/workflows/issue-labeled.yml
@@ -9,12 +9,17 @@ jobs:
if: github.repository == 'vitejs/vite'
runs-on: ubuntu-slim
permis... | [
".github/workflows/issue-labeled.yml"
] | [
{
"comment": "This workflow only runs on `issues` events (not PRs), but it still requests `pull-requests: write`. Dropping unnecessary token permissions reduces blast radius if the token is ever misused.",
"path": ".github/workflows/issue-labeled.yml",
"hunk": "@@ -9,21 +9,26 @@ jobs:\n if: github.r... | true | ||
vitejs/vite | 22,584 | comment_to_fix | ci: rewrite issue-helper with JS script | This scheduled job is only intended to operate on issues with the `needs reproduction` label, but it still requests `pull-requests: write`. If the helper is updated to skip PRs, this permission can be removed to follow least-privilege. | f94df87ff03b40b65e29bacdc04cc18c7bccaa4a | de76b899686df250779de2bfe88388750e9a84ea | diff --git a/.github/workflows/issue-close-require.yml b/.github/workflows/issue-close-require.yml
index f4fc17e5b7091f..e63d2dc61704ac 100644
--- a/.github/workflows/issue-close-require.yml
+++ b/.github/workflows/issue-close-require.yml
@@ -9,11 +9,16 @@ jobs:
if: github.repository == 'vitejs/vite'
runs-on:... | [
".github/workflows/issue-close-require.yml"
] | [
{
"comment": "This scheduled job is only intended to operate on issues with the `needs reproduction` label, but it still requests `pull-requests: write`. If the helper is updated to skip PRs, this permission can be removed to follow least-privilege.",
"path": ".github/workflows/issue-close-require.yml",
... | true | ||
vitejs/vite | 22,584 | comment_to_fix | ci: rewrite issue-helper with JS script | ```suggestion
if (!labeledAt || new Date(labeledAt) > cutoff || new Date(issue.updated_at) > cutoff) continue
```
https://github.com/vitejs/vite/pull/22584/changes#r3338046881 | f94df87ff03b40b65e29bacdc04cc18c7bccaa4a | de76b899686df250779de2bfe88388750e9a84ea | diff --git a/.github/actions/issues-helper/action.yml b/.github/actions/issues-helper/action.yml
new file mode 100644
index 00000000000000..f2aeb0937f2cc9
--- /dev/null
+++ b/.github/actions/issues-helper/action.yml
@@ -0,0 +1,168 @@
+name: Issues Helper
+description: Add labels, create comments, close stale issues, or... | [
".github/actions/issues-helper/action.yml"
] | [
{
"comment": "```suggestion\n if (!labeledAt || new Date(labeledAt) > cutoff || new Date(issue.updated_at) > cutoff) continue\n```\nhttps://github.com/vitejs/vite/pull/22584/changes#r3338046881",
"path": ".github/actions/issues-helper/action.yml",
"hunk": "@@ -0,0 +1,168 @@\n+name: Issues Hel... | true | ||
vitejs/vite | 22,572 | issue_to_patch | fix: reject windows alternate paths | reject paths containing `:` or `~`.
Fixes GHSA-fx2h-pf6j-xcff
| 8d1b0195fd186d0b3297d7cd17acff6c96797420 | b912702f088575bd49fc3cdf6c87ae2bd6ee5b59 | diff --git a/packages/vite/src/node/server/middlewares/static.ts b/packages/vite/src/node/server/middlewares/static.ts
index 122b77eafa2eed..bb093f127f5d13 100644
--- a/packages/vite/src/node/server/middlewares/static.ts
+++ b/packages/vite/src/node/server/middlewares/static.ts
@@ -286,6 +286,8 @@ export function isFil... | [
"packages/vite/src/node/server/middlewares/static.ts",
"playground/fs-serve/__tests__/commonTests.ts",
"playground/fs-serve/root/matrixTestResultPlugin.ts",
"playground/fs-serve/root/src/index.html",
"playground/fs-serve/root/vite.config-base.js",
"playground/fs-serve/root/vite.config.js",
"playground/f... | [] | diff --git a/playground/fs-serve/__tests__/commonTests.ts b/playground/fs-serve/__tests__/commonTests.ts
index 17f6199741d789..49b83904e3d051 100644
--- a/playground/fs-serve/__tests__/commonTests.ts
+++ b/playground/fs-serve/__tests__/commonTests.ts
@@ -14,6 +14,7 @@ import {
import type { Page } from 'playwright-chr... | true | |
vitejs/vite | 22,571 | issue_to_patch | fix(deps): reject UNC paths for launch-editor-middleware | https://github.com/vitejs/launch-editor/pull/138 | 8d1b0195fd186d0b3297d7cd17acff6c96797420 | 21134eb33c1b1d062dcaaf6d27cbe87259bc643d | diff --git a/packages/vite/package.json b/packages/vite/package.json
index 4afe92d0e43eef..8a8167abc84836 100644
--- a/packages/vite/package.json
+++ b/packages/vite/package.json
@@ -111,7 +111,7 @@
"etag": "^1.8.1",
"host-validation-middleware": "^0.1.4",
"http-proxy-3": "^1.23.2",
- "launch-editor-m... | [
"packages/vite/package.json",
"pnpm-lock.yaml"
] | [] | true | ||
vitejs/vite | 22,576 | issue_to_patch | fix: backport #22572, reject windows alternate paths | backport of https://github.com/vitejs/vite/pull/22572 for v6 | 8d8892bfbfb140f7fc3edecab63bddc946e9977a | fab4aa3b15c8ad117e9f18d5cb7c4fef8a721940 | diff --git a/packages/vite/src/node/server/middlewares/static.ts b/packages/vite/src/node/server/middlewares/static.ts
index b1e6393790df9c..dafd632787b9c7 100644
--- a/packages/vite/src/node/server/middlewares/static.ts
+++ b/packages/vite/src/node/server/middlewares/static.ts
@@ -266,6 +266,8 @@ export function isUri... | [
"packages/vite/src/node/server/middlewares/static.ts",
"playground/fs-serve/__tests__/commonTests.ts",
"playground/fs-serve/root/src/index.html",
"playground/fs-serve/root/vite.config-base.js",
"playground/fs-serve/root/vite.config.js",
"playground/fs-serve/root/windows83Filename.ts"
] | [] | diff --git a/playground/fs-serve/__tests__/commonTests.ts b/playground/fs-serve/__tests__/commonTests.ts
index 1493aac0271b15..5183ae280323bb 100644
--- a/playground/fs-serve/__tests__/commonTests.ts
+++ b/playground/fs-serve/__tests__/commonTests.ts
@@ -13,6 +13,7 @@ import {
import type { Page } from 'playwright-chr... | true | |
vitejs/vite | 22,575 | issue_to_patch | fix(deps): backport #22571, reject UNC paths for launch-editor-middleware | backport of https://github.com/vitejs/vite/pull/22571 for v6 | 8d8892bfbfb140f7fc3edecab63bddc946e9977a | 840ce1a02ff786aa6e53aed65d282660fe10c7a0 | diff --git a/packages/vite/LICENSE.md b/packages/vite/LICENSE.md
index af33d3a2d2209c..abcbee57e9f466 100644
--- a/packages/vite/LICENSE.md
+++ b/packages/vite/LICENSE.md
@@ -1345,7 +1345,7 @@ Repository: lydell/js-tokens
## launch-editor, launch-editor-middleware
License: MIT
By: Evan You
-Repositories: git+https:/... | [
"packages/vite/LICENSE.md",
"packages/vite/package.json",
"pnpm-lock.yaml"
] | [] | true | ||
vitejs/vite | 22,574 | issue_to_patch | fix: backport #22572, reject windows alternate paths | backport of https://github.com/vitejs/vite/pull/22572 for v7 | 957839804ad88b703f111b8935ea35930447f57f | 95f393c3d3b1a6545a7dbfac3b7358612e042a83 | diff --git a/packages/vite/src/node/server/middlewares/static.ts b/packages/vite/src/node/server/middlewares/static.ts
index 9dfc8238ec6b07..07c196498a8786 100644
--- a/packages/vite/src/node/server/middlewares/static.ts
+++ b/packages/vite/src/node/server/middlewares/static.ts
@@ -286,6 +286,8 @@ export function isFil... | [
"packages/vite/src/node/server/middlewares/static.ts",
"playground/fs-serve/__tests__/commonTests.ts",
"playground/fs-serve/root/matrixTestResultPlugin.ts",
"playground/fs-serve/root/src/index.html",
"playground/fs-serve/root/vite.config-base.js",
"playground/fs-serve/root/vite.config.js",
"playground/f... | [] | diff --git a/playground/fs-serve/__tests__/commonTests.ts b/playground/fs-serve/__tests__/commonTests.ts
index 95cbda57b1825b..9805c8b8af5ce8 100644
--- a/playground/fs-serve/__tests__/commonTests.ts
+++ b/playground/fs-serve/__tests__/commonTests.ts
@@ -14,6 +14,7 @@ import {
import type { Page } from 'playwright-chr... | true | |
vitejs/vite | 22,573 | issue_to_patch | fix(deps): backport #22571, reject UNC paths for launch-editor-middleware | backport of #22571 for v7
<!--
- What is this PR solving? Write a clear and concise description.
- Reference the issues it solves (e.g. `fixes #123`).
- What other alternatives have you explored?
- Are there any parts you think require more attention from reviewers?
Also, please make sure you do the following... | 957839804ad88b703f111b8935ea35930447f57f | 73c3582328051d51a60b2c0c21f7597ab9a7f193 | diff --git a/packages/vite/LICENSE.md b/packages/vite/LICENSE.md
index c0f6ed8846e407..c49b2ae8891a1b 100644
--- a/packages/vite/LICENSE.md
+++ b/packages/vite/LICENSE.md
@@ -1085,7 +1085,7 @@ Repository: https://github.com/lydell/js-tokens
## launch-editor, launch-editor-middleware
License: MIT
By: Evan You
-Reposi... | [
"packages/vite/LICENSE.md",
"packages/vite/package.json",
"pnpm-lock.yaml"
] | [] | true | ||
vitejs/vite | 22,562 | issue_to_patch | refactor: correct logic in `collectAllModules` function | Hi team,
While going through the source code, I came across this function:
https://github.com/vitejs/vite/blob/646dbedd2870f8ec48df0321177d8aa64bbd1575/packages/vite/src/node/config.ts#L2519
I noticed one thing about the collectAllModules implementation. Inside the for loops, we are adding the imported chunk f... | 646dbedd2870f8ec48df0321177d8aa64bbd1575 | a47ba4017a09d4475a0e7a6ec1168a4cdf315e8b | diff --git a/packages/vite/src/node/config.ts b/packages/vite/src/node/config.ts
index b3a29711cd45cd..53bdd1ca7c1b4a 100644
--- a/packages/vite/src/node/config.ts
+++ b/packages/vite/src/node/config.ts
@@ -2525,16 +2525,16 @@ function collectAllModules(
if (analyzedModules.has(fileName)) return
analyzedModules.a... | [
"packages/vite/src/node/config.ts"
] | [] | true | ||
vllm-project/vllm | 45,383 | issue_to_patch | [BugFix] Fix prompt_embeds for multimodal models | ## Purpose
Fix `prompt_embeds` for multimodal models (e.g. `Gemma3ForConditionalGeneration`).
Related to #44842.
With `LLM(..., enable_prompt_embeds=True).generate({"prompt_embeds": ...})`, a
multimodal model currently produces **crashes the
EngineCore** with a CUDA index-out-of-bounds assertion
(`vectorized... | 0d29612292c6b1e312af42ac00cf649af16a438b | f251b8dc701ae918038c8bd08b396f7bb5fe7a4e | diff --git a/vllm/config/vllm.py b/vllm/config/vllm.py
index 6122476abb8e..6889508e9465 100644
--- a/vllm/config/vllm.py
+++ b/vllm/config/vllm.py
@@ -949,6 +949,15 @@ def __post_init__(self):
"Async scheduling is not compatible with "
"disable_padded_drafter_batch=True... | [
"vllm/config/vllm.py",
"vllm/v1/worker/gpu_model_runner.py"
] | [
{
"comment": "For my own understanding, what would be necessary to be able to support Async Scheduling + prompt embeds + multimodal?",
"path": "vllm/config/vllm.py",
"hunk": "@@ -947,6 +935,15 @@ def __post_init__(self):\n \"Async scheduling is not compatible with \"\n ... | true | ||
vllm-project/vllm | 45,566 | issue_to_patch | [Perf] Use bisect for mm feature lookup in model runner v2 |
## Purpose
Port changes in #44212 to the model runner v2 encoder runner.
## Test Plan
## Test Result
---
<details>
<summary> Essential Elements of an Effective PR Description Checklist </summary>
- [ ] The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
- [ ]... | 78e7293bb157498d23780891cc4ef365a31772b6 | fa622c9b81922f911ffbe857a0cef5cb2b88be09 | diff --git a/vllm/v1/worker/gpu/mm/encoder_runner.py b/vllm/v1/worker/gpu/mm/encoder_runner.py
index 1000dbe05a80..aa636cf245fe 100644
--- a/vllm/v1/worker/gpu/mm/encoder_runner.py
+++ b/vllm/v1/worker/gpu/mm/encoder_runner.py
@@ -5,7 +5,7 @@
from vllm.model_executor.models.interfaces import SupportsMultiModal
from... | [
"vllm/v1/worker/gpu/mm/encoder_runner.py"
] | [] | true | ||
vllm-project/vllm | 44,400 | issue_to_patch | [ROCm][Perf] Enable W4A16 FlyDSL MoE | ## Purpose
This PR replaces w4a16 MoE GPTQ Triton kernel with custom FlyDSL w4a16 MoE kernel if
1. weights are int 4-bit quantized
2. group_size==32
3. arch==gfx950 (currently limited to gfx950, as the MoE was tuned only for gfx950)
4. --moe-backend=flydsl
5. lora is disabled
targeting Kimi K2.5 INT4 model. ... | 470229c37efaf69c86e8bc97482b0b1ff7551c65 | 7304e38fbb8dbbad32df2482d9e777f0a85cb85e | diff --git a/benchmarks/kernels/benchmark_flydsl_moe_w4a16.py b/benchmarks/kernels/benchmark_flydsl_moe_w4a16.py
new file mode 100644
index 000000000000..9e4f4157a8a4
--- /dev/null
+++ b/benchmarks/kernels/benchmark_flydsl_moe_w4a16.py
@@ -0,0 +1,277 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: ... | [
"benchmarks/kernels/benchmark_flydsl_moe_w4a16.py",
"tests/kernels/moe/test_flydsl_moe.py",
"vllm/config/kernel.py",
"vllm/model_executor/layers/fused_moe/configs/E=384,N=256,device_name=AMD_Instinct_MI350X,dtype=int4_w4a16,backend=flydsl.json",
"vllm/model_executor/layers/fused_moe/configs/E=384,N=256,devi... | [
{
"comment": "For MOE, we do not rely on environment flag.\n\nWe should use the python argument interface instead\n\n`--moe-backend` https://docs.vllm.ai/en/latest/cli/serve/#-moe-backend\n\nyou can name is `flydsl`",
"path": "vllm/envs.py",
"hunk": "@@ -1102,6 +1102,11 @@ def _resolve_rust_frontend_pat... | diff --git a/tests/kernels/moe/test_flydsl_moe.py b/tests/kernels/moe/test_flydsl_moe.py
new file mode 100644
index 000000000000..7c51c3691311
--- /dev/null
+++ b/tests/kernels/moe/test_flydsl_moe.py
@@ -0,0 +1,179 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM p... | true | |
vllm-project/vllm | 44,400 | comment_to_fix | [ROCm][Perf] Enable W4A16 FlyDSL MoE | Is this kernel generalised to all type of input shapes and expert parallelism configuration?
This question will help us to understand if we should set it as a default for gfx950. | 470229c37efaf69c86e8bc97482b0b1ff7551c65 | 7304e38fbb8dbbad32df2482d9e777f0a85cb85e | diff --git a/vllm/model_executor/layers/fused_moe/fused_flydsl_moe.py b/vllm/model_executor/layers/fused_moe/fused_flydsl_moe.py
new file mode 100644
index 000000000000..cf49e01e6282
--- /dev/null
+++ b/vllm/model_executor/layers/fused_moe/fused_flydsl_moe.py
@@ -0,0 +1,430 @@
+# SPDX-License-Identifier: Apache-2.0
+# ... | [
"vllm/model_executor/layers/fused_moe/fused_flydsl_moe.py"
] | [
{
"comment": "Is this kernel generalised to all type of input shapes and expert parallelism configuration?\nThis question will help us to understand if we should set it as a default for gfx950.",
"path": "vllm/model_executor/layers/fused_moe/fused_flydsl_moe.py",
"hunk": "@@ -0,0 +1,392 @@\n+# SPDX-Lice... | true | ||
vllm-project/vllm | 44,400 | comment_to_fix | [ROCm][Perf] Enable W4A16 FlyDSL MoE | @amd-asalykov please include the tuning script you used to generate these `.json` file in `benchmarks/kernels`. | 470229c37efaf69c86e8bc97482b0b1ff7551c65 | 7304e38fbb8dbbad32df2482d9e777f0a85cb85e | diff --git a/vllm/model_executor/layers/fused_moe/configs/E=384,N=256,device_name=AMD_Instinct_MI350_OAM,dtype=int4_w4a16,backend=flydsl.json b/vllm/model_executor/layers/fused_moe/configs/E=384,N=256,device_name=AMD_Instinct_MI350_OAM,dtype=int4_w4a16,backend=flydsl.json
new file mode 100644
index 000000000000..bb1a95... | [
"vllm/model_executor/layers/fused_moe/configs/E=384,N=256,device_name=AMD_Instinct_MI350_OAM,dtype=int4_w4a16,backend=flydsl.json"
] | [
{
"comment": "@amd-asalykov please include the tuning script you used to generate these `.json` file in `benchmarks/kernels`.",
"path": "vllm/model_executor/layers/fused_moe/configs/E=384,N=256,device_name=AMD_Instinct_MI350_OAM,dtype=int4_w4a16,backend=flydsl.json",
"hunk": "@@ -0,0 +1,114 @@\n+{",
... | true | ||
vllm-project/vllm | 44,400 | comment_to_fix | [ROCm][Perf] Enable W4A16 FlyDSL MoE | for the token size please add a test till `16384`. This is the typical `max-num-batched-token` size.
Offline inference `max-num-batched-token` is `16384`
Online inference `max-num-batched-token` is `8192` | 470229c37efaf69c86e8bc97482b0b1ff7551c65 | 7304e38fbb8dbbad32df2482d9e777f0a85cb85e | diff --git a/tests/kernels/moe/test_flydsl_moe.py b/tests/kernels/moe/test_flydsl_moe.py
new file mode 100644
index 000000000000..7c51c3691311
--- /dev/null
+++ b/tests/kernels/moe/test_flydsl_moe.py
@@ -0,0 +1,179 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM p... | [
"tests/kernels/moe/test_flydsl_moe.py"
] | [
{
"comment": "for the token size please add a test till `16384`. This is the typical `max-num-batched-token` size.\n\nOffline inference `max-num-batched-token` is `16384`\nOnline inference `max-num-batched-token` is `8192`",
"path": "tests/kernels/moe/test_flydsl_moe.py",
"hunk": "@@ -0,0 +1,224 @@\n+# ... | true | ||
vllm-project/vllm | 44,400 | comment_to_fix | [ROCm][Perf] Enable W4A16 FlyDSL MoE | Can these tuning files be stored in AITER instead? | 470229c37efaf69c86e8bc97482b0b1ff7551c65 | 7304e38fbb8dbbad32df2482d9e777f0a85cb85e | diff --git a/vllm/model_executor/layers/fused_moe/configs/E=384,N=256,device_name=AMD_Instinct_MI350_OAM,dtype=int4_w4a16,backend=flydsl.json b/vllm/model_executor/layers/fused_moe/configs/E=384,N=256,device_name=AMD_Instinct_MI350_OAM,dtype=int4_w4a16,backend=flydsl.json
new file mode 100644
index 000000000000..bb1a95... | [
"vllm/model_executor/layers/fused_moe/configs/E=384,N=256,device_name=AMD_Instinct_MI350_OAM,dtype=int4_w4a16,backend=flydsl.json"
] | [
{
"comment": "Can these tuning files be stored in AITER instead?",
"path": "vllm/model_executor/layers/fused_moe/configs/E=384,N=256,device_name=AMD_Instinct_MI350_OAM,dtype=int4_w4a16,backend=flydsl.json",
"hunk": "@@ -0,0 +1,114 @@\n+{",
"resolving_sha": "7304e38fbb8dbbad32df2482d9e777f0a85cb85e",... | true | ||
vllm-project/vllm | 44,400 | comment_to_fix | [ROCm][Perf] Enable W4A16 FlyDSL MoE | process_weights_after_loading runs before select_gemm_impl. By the time this branch runs, process_weights_after_loading has already done w13.view(-1) / w2.view(-1), so the packed tensors are flat 1-D blobs in the FlyDSL packed-int4 byte layout not the [E, K//8, N] int32 GPTQ layout that TritonWNA16Experts expects. The ... | 470229c37efaf69c86e8bc97482b0b1ff7551c65 | 7304e38fbb8dbbad32df2482d9e777f0a85cb85e | diff --git a/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe/compressed_tensors_moe_w4a16_flydsl.py b/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe/compressed_tensors_moe_w4a16_flydsl.py
new file mode 100644
index 000000000000..f8faddbd07bf
--- /dev/... | [
"vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe/compressed_tensors_moe_w4a16_flydsl.py"
] | [
{
"comment": "process_weights_after_loading runs before select_gemm_impl. By the time this branch runs, process_weights_after_loading has already done w13.view(-1) / w2.view(-1), so the packed tensors are flat 1-D blobs in the FlyDSL packed-int4 byte layout not the [E, K//8, N] int32 GPTQ layout that TritonWNA1... | true | ||
vllm-project/vllm | 45,136 | issue_to_patch | [XPU] Support int4 group_size=32 W4A16 MoE | ## Purpose
The XPU WNA16 fused-MoE expert backend (XPUExpertsWNA16) accepted only the catch-all int4 weight scale key kInt4Static (GroupShape(1, -1)). In the compressed-tensors WNA16 mapping, this catch-all key is assigned to every int4 symmetric scheme except group_size=32 — i.e. per-channel as well as group_size 6... | 17ee5b1ac5dd61fa89bc4321ef54b0a790a45db3 | 0cf73a65785185e4a4cf73f521e8977f6626cdb5 | diff --git a/vllm/model_executor/layers/fused_moe/experts/xpu_moe.py b/vllm/model_executor/layers/fused_moe/experts/xpu_moe.py
index 00829a0f7086..fe86e2b35ff7 100644
--- a/vllm/model_executor/layers/fused_moe/experts/xpu_moe.py
+++ b/vllm/model_executor/layers/fused_moe/experts/xpu_moe.py
@@ -19,6 +19,7 @@
kFp8St... | [
"vllm/model_executor/layers/fused_moe/experts/xpu_moe.py"
] | [] | true | ||
vllm-project/vllm | 45,487 | issue_to_patch | [Bugfix][DCP] Fix illegal memory access in DCP a2a decode under full CUDA graphs | ## Summary
`--dcp-comm-backend a2a` crashes every worker with a CUDA illegal memory access on the first request when full CUDA graphs are enabled (cudagraph_mode=FULL_AND_PIECEWISE) for DCP4 when running model `nvidia/Kimi-K2.5-NVFP4`. The DCP a2a combine drew its all-to-all send/recv buffers from the shared, growab... | ff5a30cfac59c9c753b6340a59c6d8ed668752f0 | 6df7cb5f2d5b3c98e8cccd1fe344579b2c663383 | diff --git a/vllm/v1/attention/ops/dcp_alltoall.py b/vllm/v1/attention/ops/dcp_alltoall.py
index 1469a5c754d6..5effeea5fb38 100644
--- a/vllm/v1/attention/ops/dcp_alltoall.py
+++ b/vllm/v1/attention/ops/dcp_alltoall.py
@@ -26,10 +26,6 @@
import torch.distributed as dist
from vllm.triton_utils import tl, triton
-fro... | [
"vllm/v1/attention/ops/dcp_alltoall.py"
] | [] | true | ||
vllm-project/vllm | 32,374 | issue_to_patch | [V1][Spec Decode] Add Dynamic SD | ## Why is Dynamic SD needed?
SD methods need to verify K tokens for each sequence during decoding. As BS increases, the effective BS becomes BS * K which increases the compute requirement during verification. When this BS*K goes beyond a critical BS then SD negatively impacts the TPOT. DSD helps by tuning down the K t... | 17ee5b1ac5dd61fa89bc4321ef54b0a790a45db3 | 3ce163938f4a654e8f4e9cc6d9f8fcb8f1c4f6ae | diff --git a/docs/features/speculative_decoding/README.md b/docs/features/speculative_decoding/README.md
index 58d1df9dced0..7213ef41ecd0 100644
--- a/docs/features/speculative_decoding/README.md
+++ b/docs/features/speculative_decoding/README.md
@@ -17,6 +17,7 @@ vLLM supports a variety of methods of speculative decod... | [
"docs/features/speculative_decoding/README.md",
"docs/features/speculative_decoding/dynamic_speculative_decoding.md",
"tests/v1/spec_decode/test_dynamic_sd.py",
"tests/v1/spec_decode/test_eagle.py",
"tests/v1/spec_decode/test_extract_hidden_states.py",
"tests/v1/spec_decode/test_mtp.py",
"tests/v1/spec_... | [
{
"comment": "\n\nA `ZeroDivisionError` will occur here if `larger_bs == smaller_bs`. This can happen if `batch_size` is outside the range of profiled batch sizes, or if only one batch size was profiled. This will crash the vLLM server. You should... | diff --git a/tests/v1/spec_decode/test_dynamic_sd.py b/tests/v1/spec_decode/test_dynamic_sd.py
new file mode 100644
index 000000000000..fe9f30ba25f2
--- /dev/null
+++ b/tests/v1/spec_decode/test_dynamic_sd.py
@@ -0,0 +1,218 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to t... | true | |
vllm-project/vllm | 32,374 | comment_to_fix | [V1][Spec Decode] Add Dynamic SD | 
The code updates `self.num_speculative_tokens`, but `batch_propose` (which is called by `propose`) uses `self.k` to determine the number of speculative tokens. This means the dynamically provided `optimal_num_speculative_tokens` has no effect. This is a... | 17ee5b1ac5dd61fa89bc4321ef54b0a790a45db3 | 3ce163938f4a654e8f4e9cc6d9f8fcb8f1c4f6ae | diff --git a/vllm/v1/spec_decode/ngram_proposer.py b/vllm/v1/spec_decode/ngram_proposer.py
index 53199d0ce217..e0240d0e66b8 100644
--- a/vllm/v1/spec_decode/ngram_proposer.py
+++ b/vllm/v1/spec_decode/ngram_proposer.py
@@ -55,6 +55,7 @@ def __init__(self, vllm_config: VllmConfig):
# Trigger Numba JIT compilati... | [
"vllm/v1/spec_decode/ngram_proposer.py"
] | [
{
"comment": "\n\nThe code updates `self.num_speculative_tokens`, but `batch_propose` (which is called by `propose`) uses `self.k` to determine the number of speculative tokens. This means the dynamically provided `optimal_num_speculative_tokens` ... | true | ||
vllm-project/vllm | 32,374 | comment_to_fix | [V1][Spec Decode] Add Dynamic SD | 
The code directly uses `dynamic_config_path` to open a file. This could be a path traversal vulnerability if the path is controlled by a user and not properly sanitized. An attacker could potentially read arbitrary files on the system. It is recommende... | 17ee5b1ac5dd61fa89bc4321ef54b0a790a45db3 | 3ce163938f4a654e8f4e9cc6d9f8fcb8f1c4f6ae | diff --git a/vllm/config/speculative.py b/vllm/config/speculative.py
index a4d5b1302e63..eba8653d63b6 100644
--- a/vllm/config/speculative.py
+++ b/vllm/config/speculative.py
@@ -157,6 +157,14 @@ class SpeculativeConfig:
target_parallel_config: SkipValidation[ParallelConfig] = None # type: ignore
"""The para... | [
"vllm/config/speculative.py"
] | [
{
"comment": "\n\nThe code directly uses `dynamic_config_path` to open a file. This could be a path traversal vulnerability if the path is controlled by a user and not properly sanitized. An attacker could potentially read arbitrary files on the ... | true | ||
vllm-project/vllm | 32,374 | comment_to_fix | [V1][Spec Decode] Add Dynamic SD | ### NgramProposer initialization missing required argument after signature change
**High Severity**
<!-- DESCRIPTION START -->
The `propose` method signature was changed to add `optimal_num_speculative_tokens` as the first parameter, but the call in `__init__` (lines 56-62) was not updated. This causes all arguments ... | 17ee5b1ac5dd61fa89bc4321ef54b0a790a45db3 | 3ce163938f4a654e8f4e9cc6d9f8fcb8f1c4f6ae | diff --git a/vllm/v1/spec_decode/ngram_proposer.py b/vllm/v1/spec_decode/ngram_proposer.py
index 53199d0ce217..e0240d0e66b8 100644
--- a/vllm/v1/spec_decode/ngram_proposer.py
+++ b/vllm/v1/spec_decode/ngram_proposer.py
@@ -55,6 +55,7 @@ def __init__(self, vllm_config: VllmConfig):
# Trigger Numba JIT compilati... | [
"vllm/v1/spec_decode/ngram_proposer.py"
] | [
{
"comment": "### NgramProposer initialization missing required argument after signature change\n\n**High Severity**\n\n<!-- DESCRIPTION START -->\nThe `propose` method signature was changed to add `optimal_num_speculative_tokens` as the first parameter, but the call in `__init__` (lines 56-62) was not updated.... | true | ||
vllm-project/vllm | 32,374 | comment_to_fix | [V1][Spec Decode] Add Dynamic SD | ### Dynamic SD adjustment uses wrong attribute in NgramProposer
**Medium Severity**
<!-- DESCRIPTION START -->
The `propose` method sets `self.num_speculative_tokens` when `optimal_num_speculative_tokens` is provided, but `NgramProposer` uses `self.k` (set in `__init__` at line 24) for the actual number of speculativ... | 17ee5b1ac5dd61fa89bc4321ef54b0a790a45db3 | 3ce163938f4a654e8f4e9cc6d9f8fcb8f1c4f6ae | diff --git a/vllm/v1/spec_decode/ngram_proposer.py b/vllm/v1/spec_decode/ngram_proposer.py
index 53199d0ce217..e0240d0e66b8 100644
--- a/vllm/v1/spec_decode/ngram_proposer.py
+++ b/vllm/v1/spec_decode/ngram_proposer.py
@@ -55,6 +55,7 @@ def __init__(self, vllm_config: VllmConfig):
# Trigger Numba JIT compilati... | [
"vllm/v1/spec_decode/ngram_proposer.py"
] | [
{
"comment": "### Dynamic SD adjustment uses wrong attribute in NgramProposer\n\n**Medium Severity**\n\n<!-- DESCRIPTION START -->\nThe `propose` method sets `self.num_speculative_tokens` when `optimal_num_speculative_tokens` is provided, but `NgramProposer` uses `self.k` (set in `__init__` at line 24) for the ... | true | ||
vllm-project/vllm | 32,374 | comment_to_fix | [V1][Spec Decode] Add Dynamic SD | Please use the same config format as all the other config classes
```suggestion
@config
class DynamicSpeculativeConfig:
``` | 17ee5b1ac5dd61fa89bc4321ef54b0a790a45db3 | 3ce163938f4a654e8f4e9cc6d9f8fcb8f1c4f6ae | diff --git a/vllm/config/speculative.py b/vllm/config/speculative.py
index a4d5b1302e63..eba8653d63b6 100644
--- a/vllm/config/speculative.py
+++ b/vllm/config/speculative.py
@@ -157,6 +157,14 @@ class SpeculativeConfig:
target_parallel_config: SkipValidation[ParallelConfig] = None # type: ignore
"""The para... | [
"vllm/config/speculative.py"
] | [
{
"comment": "Please use the same config format as all the other config classes\n\n```suggestion\n@config\nclass DynamicSpeculativeConfig:\n```",
"path": "vllm/config/speculative.py",
"hunk": "@@ -53,6 +53,42 @@\n ]\n \n \n+class DynamicSpeculativeConfig(BaseModel):",
"resolving_sha": "3ce163938f4a6... | true | ||
vllm-project/vllm | 32,374 | comment_to_fix | [V1][Spec Decode] Add Dynamic SD | Docstrings go under what they are documenting | 17ee5b1ac5dd61fa89bc4321ef54b0a790a45db3 | 3ce163938f4a654e8f4e9cc6d9f8fcb8f1c4f6ae | diff --git a/vllm/config/speculative.py b/vllm/config/speculative.py
index a4d5b1302e63..eba8653d63b6 100644
--- a/vllm/config/speculative.py
+++ b/vllm/config/speculative.py
@@ -157,6 +157,14 @@ class SpeculativeConfig:
target_parallel_config: SkipValidation[ParallelConfig] = None # type: ignore
"""The para... | [
"vllm/config/speculative.py"
] | [
{
"comment": "Docstrings go under what they are documenting",
"path": "vllm/config/speculative.py",
"hunk": "@@ -53,6 +53,42 @@\n ]\n \n \n+class DynamicSpeculativeConfig(BaseModel):\n+ # \"\"\"A mapping from batch size to optimal number of drafts to use for that\n+ # batch size. This is used to d... | true | ||
vllm-project/vllm | 32,374 | comment_to_fix | [V1][Spec Decode] Add Dynamic SD | This should be specified as a field in the dataclass, not dynamically added in post init | 17ee5b1ac5dd61fa89bc4321ef54b0a790a45db3 | 3ce163938f4a654e8f4e9cc6d9f8fcb8f1c4f6ae | diff --git a/vllm/config/speculative.py b/vllm/config/speculative.py
index a4d5b1302e63..eba8653d63b6 100644
--- a/vllm/config/speculative.py
+++ b/vllm/config/speculative.py
@@ -157,6 +157,14 @@ class SpeculativeConfig:
target_parallel_config: SkipValidation[ParallelConfig] = None # type: ignore
"""The para... | [
"vllm/config/speculative.py"
] | [
{
"comment": "This should be specified as a field in the dataclass, not dynamically added in post init",
"path": "vllm/config/speculative.py",
"hunk": "@@ -601,6 +640,18 @@ def __post_init__(self):\n self.target_parallel_config, self.draft_tensor_parallel_size\n ... | true | ||
vllm-project/vllm | 32,374 | comment_to_fix | [V1][Spec Decode] Add Dynamic SD | Surely there's a pytest built-in for this? | 17ee5b1ac5dd61fa89bc4321ef54b0a790a45db3 | 3ce163938f4a654e8f4e9cc6d9f8fcb8f1c4f6ae | diff --git a/tests/v1/spec_decode/test_dynamic_sd.py b/tests/v1/spec_decode/test_dynamic_sd.py
new file mode 100644
index 000000000000..fe9f30ba25f2
--- /dev/null
+++ b/tests/v1/spec_decode/test_dynamic_sd.py
@@ -0,0 +1,218 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to t... | [
"tests/v1/spec_decode/test_dynamic_sd.py"
] | [
{
"comment": "Surely there's a pytest built-in for this?",
"path": "tests/v1/spec_decode/test_dynamic_sd.py",
"hunk": "@@ -0,0 +1,265 @@\n+# SPDX-License-Identifier: Apache-2.0\n+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project\n+\"\"\"Regression tests for the Dynamic SD batch-size s... | true | ||
vllm-project/vllm | 32,374 | comment_to_fix | [V1][Spec Decode] Add Dynamic SD | This seems like a pointless abstraction if it's just a stand-in for `num_speculative_tokens_per_batch_size: dict[str, int]` that doesn't do any additional validation | 17ee5b1ac5dd61fa89bc4321ef54b0a790a45db3 | 3ce163938f4a654e8f4e9cc6d9f8fcb8f1c4f6ae | diff --git a/vllm/config/speculative.py b/vllm/config/speculative.py
index a4d5b1302e63..eba8653d63b6 100644
--- a/vllm/config/speculative.py
+++ b/vllm/config/speculative.py
@@ -157,6 +157,14 @@ class SpeculativeConfig:
target_parallel_config: SkipValidation[ParallelConfig] = None # type: ignore
"""The para... | [
"vllm/config/speculative.py"
] | [
{
"comment": "This seems like a pointless abstraction if it's just a stand-in for `num_speculative_tokens_per_batch_size: dict[str, int]` that doesn't do any additional validation",
"path": "vllm/config/speculative.py",
"hunk": "@@ -70,6 +70,21 @@\n DraftSampleMethod = Literal[\"greedy\", \"probabilisti... | true | ||
vllm-project/vllm | 32,374 | comment_to_fix | [V1][Spec Decode] Add Dynamic SD | linting issue? | 17ee5b1ac5dd61fa89bc4321ef54b0a790a45db3 | 3ce163938f4a654e8f4e9cc6d9f8fcb8f1c4f6ae | diff --git a/vllm/config/speculative.py b/vllm/config/speculative.py
index a4d5b1302e63..eba8653d63b6 100644
--- a/vllm/config/speculative.py
+++ b/vllm/config/speculative.py
@@ -157,6 +157,14 @@ class SpeculativeConfig:
target_parallel_config: SkipValidation[ParallelConfig] = None # type: ignore
"""The para... | [
"vllm/config/speculative.py"
] | [
{
"comment": "linting issue?",
"path": "vllm/config/speculative.py",
"hunk": "@@ -792,6 +811,7 @@ def __post_init__(self):\n self.target_parallel_config, self.draft_tensor_parallel_size\n )\n )\n+",
"resolving_sha": "3ce163938f4a654e8f4e9cc... | true | ||
vllm-project/vllm | 32,374 | comment_to_fix | [V1][Spec Decode] Add Dynamic SD | ```suggestion
| Dynamic Speculative Decoding | High gain | Higher than base SD method | Useful for RL or workload with fluctuating QPS |
``` | 17ee5b1ac5dd61fa89bc4321ef54b0a790a45db3 | 3ce163938f4a654e8f4e9cc6d9f8fcb8f1c4f6ae | diff --git a/docs/features/speculative_decoding/README.md b/docs/features/speculative_decoding/README.md
index 58d1df9dced0..7213ef41ecd0 100644
--- a/docs/features/speculative_decoding/README.md
+++ b/docs/features/speculative_decoding/README.md
@@ -17,6 +17,7 @@ vLLM supports a variety of methods of speculative decod... | [
"docs/features/speculative_decoding/README.md"
] | [
{
"comment": "```suggestion\n| Dynamic Speculative Decoding | High gain | Higher than base SD method | Useful for RL or workload with fluctuating QPS |\n```",
"path": "docs/features/speculative_decoding/README.md",
"hunk": "@@ -33,6 +34,7 @@ depend on your model family, traffic pattern, hardware, and sa... | true | ||
vllm-project/vllm | 45,179 | issue_to_patch | [Core] Release cached device memory under pressure on UMA GPUs during weight loading | ## Purpose
Alternative to #44502. Weight repacking in `process_weights_after_loading` leaves freed checkpoint-format tensors in the caching allocator; on integrated (UMA) devices like DGX Spark the reserved cache starves the OS, which can thrash before an OOM retry would release it.
Differences from #44502:
- One cal... | 40e065e86a91b312f5b4b20921cde86fa0e577e3 | cb095d4d996ff554eb9889ddb6c596c0583ad105 | diff --git a/vllm/model_executor/model_loader/utils.py b/vllm/model_executor/model_loader/utils.py
index 2a5f746d783e..fc279c7e9c78 100644
--- a/vllm/model_executor/model_loader/utils.py
+++ b/vllm/model_executor/model_loader/utils.py
@@ -30,6 +30,7 @@
)
from vllm.model_executor.models.interfaces import SupportsQuant... | [
"vllm/model_executor/model_loader/utils.py",
"vllm/utils/mem_utils.py"
] | [] | true | ||
vllm-project/vllm | 44,795 | issue_to_patch | [Bug] nightly Docker images crash with ImportError: AnthropicOutputConfig since May 28
### Your current environment
Collecting environment information...
==============================
System Info
==============================
OS : Ubuntu 22.04.5 LTS (x86_64)
GCC version ... | [Bugfix] nightly Docker images crash with ImportError: AnthropicOutputConfig since May 28 | Fixes #44759
The nightly images have been broken since May 28 with this at startup:
ImportError: cannot import name 'AnthropicOutputConfig'
from 'vllm.entrypoints.anthropic.protocol'
The source on main is fine — both `protocol.py` and `serving.py` were updated
together in #42396. The issue is in the ... | 71b961dd356a399150d25738c175c71859aa1301 | dd119eab68c5846f7917c879ad52529a42534ae3 | diff --git a/docker/Dockerfile b/docker/Dockerfile
index d03da7bcc370..7a3cc71d339b 100644
--- a/docker/Dockerfile
+++ b/docker/Dockerfile
@@ -548,9 +548,17 @@ RUN --mount=type=cache,target=/opt/uv/cache \
fi && \
python3 setup.py bdist_wheel --dist-dir=dist --py-limited-api=cp38
+# Record the wheel checksu... | [
"docker/Dockerfile",
"tests/entrypoints/anthropic/test_protocol_exports.py"
] | [
{
"comment": "The loop assertion is effectively redundant: if any symbol is missing, the module import at lines 9–21 will fail during test collection (before the test runs), and `assert export is not None` doesn’t add meaningful signal. Consider importing the protocol module inside the test and validating names... | diff --git a/tests/entrypoints/anthropic/test_protocol_exports.py b/tests/entrypoints/anthropic/test_protocol_exports.py
new file mode 100644
index 000000000000..466f40e3ccf5
--- /dev/null
+++ b/tests/entrypoints/anthropic/test_protocol_exports.py
@@ -0,0 +1,50 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopy... | true |
vllm-project/vllm | 45,277 | issue_to_patch | [Build] Fix CUDA arch build coverage gaps | ## Purpose
This PR fixes several CUDA architecture build/dispatch gaps found while auditing
the CUDA arch matrix in #45260.
The main theme is making top-level arch requests, per-component CMake arch
filters, and runtime support checks agree with each other. Before this change,
some paths either carried mislead... | 39dee1114a2cd183a9fb72b561808b385b6c9daa | e57d3b78ef6ec5d8e4206166bd51aedcb2130750 | diff --git a/.buildkite/release-pipeline.yaml b/.buildkite/release-pipeline.yaml
index b31404bca154..897c98145341 100644
--- a/.buildkite/release-pipeline.yaml
+++ b/.buildkite/release-pipeline.yaml
@@ -1,12 +1,25 @@
# CUDA architecture lists — following PyTorch RELEASE.md
# (https://github.com/pytorch/pytorch/blob/m... | [
".buildkite/release-pipeline.yaml",
".github/workflows/scripts/build.sh",
"CMakeLists.txt",
"cmake/external_projects/qutlass.cmake",
"cmake/utils.cmake",
"csrc/libtorch_stable/cuda_vec_utils.cuh",
"csrc/libtorch_stable/moe/dsv3_router_gemm_entry.cu",
"csrc/libtorch_stable/quantization/fp4/mxfp4_expert... | [
{
"comment": "**<sub><sub></sub></sub> Disable MXFP4 dispatch when the quant op is skipped**\n\nFor CUDA 12.8 SM100 builds this branch still builds `mxfp4_blockwise_moe_kernel.cu` and advertises the Cutlass MXFP4 backend via `ENABLE_NVFP4_SM100`, bu... | true | ||
vllm-project/vllm | 45,277 | comment_to_fix | [Build] Fix CUDA arch build coverage gaps | **<sub><sub></sub></sub> Disable MXFP4 dispatch when the quant op is skipped**
For CUDA 12.8 SM100 builds this branch still builds `mxfp4_blockwise_moe_kernel.cu` and advertises the Cutlass MXFP4 backend via `ENABLE_NVFP4_SM100`, but it no longer registers... | 39dee1114a2cd183a9fb72b561808b385b6c9daa | e57d3b78ef6ec5d8e4206166bd51aedcb2130750 | diff --git a/CMakeLists.txt b/CMakeLists.txt
index c03360a5d4e3..9d109ee954ce 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -190,6 +190,11 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
# the set of architectures we want to compile for and remove the from the
# CMAKE_CUDA_FLAGS so that they are not applied globally.
... | [
"CMakeLists.txt"
] | [
{
"comment": "**<sub><sub></sub></sub> Disable MXFP4 dispatch when the quant op is skipped**\n\nFor CUDA 12.8 SM100 builds this branch still builds `mxfp4_blockwise_moe_kernel.cu` and advertises the Cutlass MXFP4 backend via `ENABLE_NVFP4_SM100`, bu... | true | ||
vllm-project/vllm | 45,277 | comment_to_fix | [Build] Fix CUDA arch build coverage gaps | `MARLIN_OTHER_ARCHS` is used to decide whether to build any Marlin sources, but the new intersection list ("7.5;8.0") becomes empty for SM12x-only builds (e.g. TORCH_CUDA_ARCH_LIST="12.0"), causing the whole Marlin block to be skipped even though Marlin explicitly supports SM12x via `12.0f` / `12.0a;12.1a`. Include SM1... | 39dee1114a2cd183a9fb72b561808b385b6c9daa | e57d3b78ef6ec5d8e4206166bd51aedcb2130750 | diff --git a/CMakeLists.txt b/CMakeLists.txt
index c03360a5d4e3..9d109ee954ce 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -190,6 +190,11 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
# the set of architectures we want to compile for and remove the from the
# CMAKE_CUDA_FLAGS so that they are not applied globally.
... | [
"CMakeLists.txt"
] | [
{
"comment": "`MARLIN_OTHER_ARCHS` is used to decide whether to build any Marlin sources, but the new intersection list (\"7.5;8.0\") becomes empty for SM12x-only builds (e.g. TORCH_CUDA_ARCH_LIST=\"12.0\"), causing the whole Marlin block to be skipped even though Marlin explicitly supports SM12x via `12.0f` / ... | true | ||
vllm-project/vllm | 45,277 | comment_to_fix | [Build] Fix CUDA arch build coverage gaps | Removing `+PTX` from `HADACORE_ARCHS` means Hadacore is no longer built when users compile only for newer archs (e.g. TORCH_CUDA_ARCH_LIST="10.0" or "12.0"). Since Hadamard schemes call `ops.hadacore_transform` without checking availability, this can lead to runtime failures in such builds. Add SM10x/SM12x archs here (... | 39dee1114a2cd183a9fb72b561808b385b6c9daa | e57d3b78ef6ec5d8e4206166bd51aedcb2130750 | diff --git a/CMakeLists.txt b/CMakeLists.txt
index c03360a5d4e3..9d109ee954ce 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -190,6 +190,11 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
# the set of architectures we want to compile for and remove the from the
# CMAKE_CUDA_FLAGS so that they are not applied globally.
... | [
"CMakeLists.txt"
] | [
{
"comment": "Removing `+PTX` from `HADACORE_ARCHS` means Hadacore is no longer built when users compile only for newer archs (e.g. TORCH_CUDA_ARCH_LIST=\"10.0\" or \"12.0\"). Since Hadamard schemes call `ops.hadacore_transform` without checking availability, this can lead to runtime failures in such builds. Ad... | true | ||
vllm-project/vllm | 45,277 | comment_to_fix | [Build] Fix CUDA arch build coverage gaps | `MARLIN_MOE_OTHER_ARCHS` is used to gate the entire Marlin MoE generation/build, but the new intersection list ("7.5;8.0") becomes empty for SM12x-only builds (e.g. TORCH_CUDA_ARCH_LIST="12.0"), skipping Marlin MoE even though `MARLIN_MOE_ARCHS` / `MARLIN_MOE_FP8_ARCHS` include SM12x. Include SM12x in `MARLIN_MOE_OTHER... | 39dee1114a2cd183a9fb72b561808b385b6c9daa | e57d3b78ef6ec5d8e4206166bd51aedcb2130750 | diff --git a/CMakeLists.txt b/CMakeLists.txt
index c03360a5d4e3..9d109ee954ce 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -190,6 +190,11 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
# the set of architectures we want to compile for and remove the from the
# CMAKE_CUDA_FLAGS so that they are not applied globally.
... | [
"CMakeLists.txt"
] | [
{
"comment": "`MARLIN_MOE_OTHER_ARCHS` is used to gate the entire Marlin MoE generation/build, but the new intersection list (\"7.5;8.0\") becomes empty for SM12x-only builds (e.g. TORCH_CUDA_ARCH_LIST=\"12.0\"), skipping Marlin MoE even though `MARLIN_MOE_ARCHS` / `MARLIN_MOE_FP8_ARCHS` include SM12x. Include ... | true | ||
vllm-project/vllm | 45,277 | comment_to_fix | [Build] Fix CUDA arch build coverage gaps | The `namespace vllm {` opens here at file scope, but its matching `} // namespace vllm` (~line 319) is inside the `#if VLLM_MXFP4_EXPERTS_QUANT_SUPPORTED` block. On a CUDA 12.8 build (`VLLM_MXFP4_EXPERTS_QUANT_SUPPORTED == 0`), the closing brace is preprocessed away while the opening `{` survives, and `nvcc` fails:
`... | 39dee1114a2cd183a9fb72b561808b385b6c9daa | e57d3b78ef6ec5d8e4206166bd51aedcb2130750 | diff --git a/csrc/libtorch_stable/quantization/fp4/mxfp4_experts_quant.cu b/csrc/libtorch_stable/quantization/fp4/mxfp4_experts_quant.cu
index 062f6018653c..20f024bcef51 100644
--- a/csrc/libtorch_stable/quantization/fp4/mxfp4_experts_quant.cu
+++ b/csrc/libtorch_stable/quantization/fp4/mxfp4_experts_quant.cu
@@ -27,15... | [
"csrc/libtorch_stable/quantization/fp4/mxfp4_experts_quant.cu"
] | [
{
"comment": "The `namespace vllm {` opens here at file scope, but its matching `} // namespace vllm` (~line 319) is inside the `#if VLLM_MXFP4_EXPERTS_QUANT_SUPPORTED` block. On a CUDA 12.8 build (`VLLM_MXFP4_EXPERTS_QUANT_SUPPORTED == 0`), the closing brace is preprocessed away while the opening `{` survives... | true | ||
vllm-project/vllm | 45,277 | comment_to_fix | [Build] Fix CUDA arch build coverage gaps | When both SM100 and SM120 arch families are present in CUDA_ARCHS, this logic selects only the SM100 arch list for QuTLASS. The `_qutlass_C` ops are called from Python without any device-capability gating (e.g., `fusedQuantizeMx` / `matmul_mxf4_bf16_tn`), and existing tests explicitly run QuTLASS on SM120 devices. Buil... | 39dee1114a2cd183a9fb72b561808b385b6c9daa | e57d3b78ef6ec5d8e4206166bd51aedcb2130750 | diff --git a/cmake/external_projects/qutlass.cmake b/cmake/external_projects/qutlass.cmake
index 273fe754bed1..66c001919b03 100644
--- a/cmake/external_projects/qutlass.cmake
+++ b/cmake/external_projects/qutlass.cmake
@@ -32,21 +32,33 @@ endif()
message(STATUS "[QUTLASS] QuTLASS is available at ${qutlass_SOURCE_DIR}"... | [
"cmake/external_projects/qutlass.cmake"
] | [
{
"comment": "When both SM100 and SM120 arch families are present in CUDA_ARCHS, this logic selects only the SM100 arch list for QuTLASS. The `_qutlass_C` ops are called from Python without any device-capability gating (e.g., `fusedQuantizeMx` / `matmul_mxf4_bf16_tn`), and existing tests explicitly run QuTLASS ... | true | ||
vllm-project/vllm | 45,546 | issue_to_patch | [Bug Fix] [MiniMax-M3] Implement EAGLE3 support on the AMD MiniMax M3 | ## Overview Problem
#fix https://github.com/vllm-project/vllm/issues/45538
hi @hongxiayang @youkaichao
+viz @andyluo7 @chunfangamd
Speculative decoding with EAGLE3 (e.g. an `Inferact/MiniMax-M3-EAGLE3` draft head) works for MiniMax-M3 on CUDA but fails at engine init on ROCm:
```
RuntimeError: Model d... | 0c6d468d5c470d2797a1a86a4d079e589d25b42f | 853eb3ef25b36d0cb35b33b69cc234bc2072a768 | diff --git a/vllm/models/minimax_m3/amd/model.py b/vllm/models/minimax_m3/amd/model.py
index 27650c8e60c2..b80d3b8b3b8c 100644
--- a/vllm/models/minimax_m3/amd/model.py
+++ b/vllm/models/minimax_m3/amd/model.py
@@ -60,7 +60,9 @@
maybe_remap_kv_scale_name,
)
from vllm.model_executor.models.interfaces import (
+ ... | [
"vllm/models/minimax_m3/amd/model.py"
] | [] | true | ||
vllm-project/vllm | 45,442 | issue_to_patch | [Core] Simplify MRV2 async output handling | MRV2 currently resolves `ModelRunnerOutput` from `AsyncModelRunnerOutput` depending on whether async scheduling is enabled.
However the executors already handle the output type appropriately depending on what is required from the upstream caller, so it's simpler for the model runner to not have to care about this. | cf567cbc71a467d8479411062917e9190ee11376 | a16f6ddbe6598ff265e2e56e3255e3a5f80e2367 | diff --git a/vllm/v1/executor/multiproc_executor.py b/vllm/v1/executor/multiproc_executor.py
index b0100c3d66ae..7bc81118e6b5 100644
--- a/vllm/v1/executor/multiproc_executor.py
+++ b/vllm/v1/executor/multiproc_executor.py
@@ -396,9 +396,7 @@ def get_response():
return responses[0] if output_rank is not No... | [
"vllm/v1/executor/multiproc_executor.py",
"vllm/v1/executor/uniproc_executor.py",
"vllm/v1/worker/gpu/model_runner.py"
] | [] | true | ||
vllm-project/vllm | 43,409 | issue_to_patch | [CPU] Support CPU W4A16 INT4 MoE |
## Purpose
Support cpu AWQ, GPTQ & compressed-tensor W4A16 INT4 MoE for Intel CPUs.
## Test Plan
python -m pytest tests/kernels/moe/test_cpu_quant_fused_moe.py -v
## Test Result
**1. AWQ W4A16 INT4 MoE**
python examples/basic/offline_inference/generate.py --model QuixiAI/Qwen3-30B-A3B-AWQ --max-mode... | 4bc83323f2ea8e85c87ae5fb5ff2d792a8f61f9d | 96535b255c5e430c7faa3213b6038ccf79743f91 | diff --git a/.buildkite/hardware_tests/cpu.yaml b/.buildkite/hardware_tests/cpu.yaml
index 3db49d579e37..a064e53ebed7 100644
--- a/.buildkite/hardware_tests/cpu.yaml
+++ b/.buildkite/hardware_tests/cpu.yaml
@@ -91,7 +91,7 @@ steps:
- tests/quantization/test_cpu_wna16.py
commands:
- |
- bash .buildkite/s... | [
".buildkite/hardware_tests/cpu.yaml",
"tests/kernels/moe/test_cpu_quant_fused_moe.py",
"tests/quantization/test_cpu_wna16.py",
"vllm/model_executor/layers/fused_moe/experts/cpu_moe.py",
"vllm/model_executor/layers/fused_moe/oracle/int_wna16.py",
"vllm/model_executor/layers/quantization/auto_gptq.py",
"v... | [
{
"comment": "\n\nThe assertion strictly requires `CPUQuantAlgo.GPTQ`, but the docstring for `prepare_int4_moe_layer_for_cpu` (line 370) suggests that `CPUQuantAlgo.AWQ` is also a valid option. If AWQ is not supported by this specific path, the d... | diff --git a/tests/kernels/moe/test_cpu_quant_fused_moe.py b/tests/kernels/moe/test_cpu_quant_fused_moe.py
index f8967b199226..d8c1b9f2cb61 100644
--- a/tests/kernels/moe/test_cpu_quant_fused_moe.py
+++ b/tests/kernels/moe/test_cpu_quant_fused_moe.py
@@ -496,5 +496,258 @@ def test_mxfp4_cpu_fused_moe_bias_swiglu(M, N, ... | true | |
vllm-project/vllm | 43,409 | comment_to_fix | [CPU] Support CPU W4A16 INT4 MoE | 
The assertion strictly requires `CPUQuantAlgo.GPTQ`, but the docstring for `prepare_int4_moe_layer_for_cpu` (line 370) suggests that `CPUQuantAlgo.AWQ` is also a valid option. If AWQ is not supported by this specific path, the docstring should be updat... | 4bc83323f2ea8e85c87ae5fb5ff2d792a8f61f9d | 96535b255c5e430c7faa3213b6038ccf79743f91 | diff --git a/vllm/model_executor/layers/fused_moe/experts/cpu_moe.py b/vllm/model_executor/layers/fused_moe/experts/cpu_moe.py
index 84740fc05703..11ed775f28e5 100644
--- a/vllm/model_executor/layers/fused_moe/experts/cpu_moe.py
+++ b/vllm/model_executor/layers/fused_moe/experts/cpu_moe.py
@@ -5,7 +5,12 @@
import torc... | [
"vllm/model_executor/layers/fused_moe/experts/cpu_moe.py"
] | [
{
"comment": "\n\nThe assertion strictly requires `CPUQuantAlgo.GPTQ`, but the docstring for `prepare_int4_moe_layer_for_cpu` (line 370) suggests that `CPUQuantAlgo.AWQ` is also a valid option. If AWQ is not supported by this specific path, the d... | true | ||
vllm-project/vllm | 43,409 | comment_to_fix | [CPU] Support CPU W4A16 INT4 MoE | 
These instance variables are redundant as they are already accessible via the properties `self.w1_scale`, `self.w2_scale`, `self.w1_zp`, and `self.w2_zp` inherited from the base class `FusedMoEExperts`. Removing them improves maintainability and ensure... | 4bc83323f2ea8e85c87ae5fb5ff2d792a8f61f9d | 96535b255c5e430c7faa3213b6038ccf79743f91 | diff --git a/vllm/model_executor/layers/fused_moe/experts/cpu_moe.py b/vllm/model_executor/layers/fused_moe/experts/cpu_moe.py
index 84740fc05703..11ed775f28e5 100644
--- a/vllm/model_executor/layers/fused_moe/experts/cpu_moe.py
+++ b/vllm/model_executor/layers/fused_moe/experts/cpu_moe.py
@@ -5,7 +5,12 @@
import torc... | [
"vllm/model_executor/layers/fused_moe/experts/cpu_moe.py"
] | [
{
"comment": "\n\nThese instance variables are redundant as they are already accessible via the properties `self.w1_scale`, `self.w2_scale`, `self.w1_zp`, and `self.w2_zp` inherited from the base class `FusedMoEExperts`. Removing them improves ma... | true | ||
vllm-project/vllm | 43,409 | comment_to_fix | [CPU] Support CPU W4A16 INT4 MoE | 
Use the inherited properties `self.w1_scale`, `self.w2_scale`, `self.w1_zp`, and `self.w2_zp` instead of the redundant local variables for better consistency and maintainability.
```suggestion
self.w1_scale,
self.w2_scale,
... | 4bc83323f2ea8e85c87ae5fb5ff2d792a8f61f9d | 96535b255c5e430c7faa3213b6038ccf79743f91 | diff --git a/vllm/model_executor/layers/fused_moe/experts/cpu_moe.py b/vllm/model_executor/layers/fused_moe/experts/cpu_moe.py
index 84740fc05703..11ed775f28e5 100644
--- a/vllm/model_executor/layers/fused_moe/experts/cpu_moe.py
+++ b/vllm/model_executor/layers/fused_moe/experts/cpu_moe.py
@@ -5,7 +5,12 @@
import torc... | [
"vllm/model_executor/layers/fused_moe/experts/cpu_moe.py"
] | [
{
"comment": "\n\nUse the inherited properties `self.w1_scale`, `self.w2_scale`, `self.w1_zp`, and `self.w2_zp` instead of the redundant local variables for better consistency and maintainability.\n\n```suggestion\n self.w1_scale,\n ... | true | ||
vllm-project/vllm | 43,409 | comment_to_fix | [CPU] Support CPU W4A16 INT4 MoE | Perhaps the process procedure can be moved to ```cpu_moe.py```, likes other backends' implementation, | 4bc83323f2ea8e85c87ae5fb5ff2d792a8f61f9d | 96535b255c5e430c7faa3213b6038ccf79743f91 | diff --git a/vllm/model_executor/layers/fused_moe/oracle/int_wna16.py b/vllm/model_executor/layers/fused_moe/oracle/int_wna16.py
index 6ad60d62e97d..8de6269e2e9e 100644
--- a/vllm/model_executor/layers/fused_moe/oracle/int_wna16.py
+++ b/vllm/model_executor/layers/fused_moe/oracle/int_wna16.py
@@ -45,6 +45,7 @@
class ... | [
"vllm/model_executor/layers/fused_moe/oracle/int_wna16.py"
] | [
{
"comment": "Perhaps the process procedure can be moved to ```cpu_moe.py```, likes other backends' implementation,",
"path": "vllm/model_executor/layers/fused_moe/oracle/int_wna16.py",
"hunk": "@@ -857,6 +870,85 @@ def convert_to_wna16_moe_kernel_format(\n w13_bias,\n w2_bias,\n... | true | ||
vllm-project/vllm | 43,409 | comment_to_fix | [CPU] Support CPU W4A16 INT4 MoE | Should raise error for GPTQ with ```desc_act=True``` | 4bc83323f2ea8e85c87ae5fb5ff2d792a8f61f9d | 96535b255c5e430c7faa3213b6038ccf79743f91 | diff --git a/vllm/model_executor/layers/fused_moe/oracle/int_wna16.py b/vllm/model_executor/layers/fused_moe/oracle/int_wna16.py
index 6ad60d62e97d..8de6269e2e9e 100644
--- a/vllm/model_executor/layers/fused_moe/oracle/int_wna16.py
+++ b/vllm/model_executor/layers/fused_moe/oracle/int_wna16.py
@@ -45,6 +45,7 @@
class ... | [
"vllm/model_executor/layers/fused_moe/oracle/int_wna16.py"
] | [
{
"comment": "Should raise error for GPTQ with ```desc_act=True```",
"path": "vllm/model_executor/layers/fused_moe/oracle/int_wna16.py",
"hunk": "@@ -857,6 +870,85 @@ def convert_to_wna16_moe_kernel_format(\n w13_bias,\n w2_bias,\n )\n+ elif backend == WNA16MoEBackend.... | true | ||
vllm-project/vllm | 43,409 | comment_to_fix | [CPU] Support CPU W4A16 INT4 MoE | bias also needs to be converted to ```float32```, if they are not ```None```. | 4bc83323f2ea8e85c87ae5fb5ff2d792a8f61f9d | 96535b255c5e430c7faa3213b6038ccf79743f91 | diff --git a/vllm/model_executor/layers/fused_moe/oracle/int_wna16.py b/vllm/model_executor/layers/fused_moe/oracle/int_wna16.py
index 6ad60d62e97d..8de6269e2e9e 100644
--- a/vllm/model_executor/layers/fused_moe/oracle/int_wna16.py
+++ b/vllm/model_executor/layers/fused_moe/oracle/int_wna16.py
@@ -45,6 +45,7 @@
class ... | [
"vllm/model_executor/layers/fused_moe/oracle/int_wna16.py"
] | [
{
"comment": "bias also needs to be converted to ```float32```, if they are not ```None```.",
"path": "vllm/model_executor/layers/fused_moe/oracle/int_wna16.py",
"hunk": "@@ -857,6 +870,85 @@ def convert_to_wna16_moe_kernel_format(\n w13_bias,\n w2_bias,\n )\n+ elif ba... | true | ||
vllm-project/vllm | 43,409 | comment_to_fix | [CPU] Support CPU W4A16 INT4 MoE | In current MoE backend abstraction, ```FusedMoEExpertsModular``` requires workspace allocation and ```FusedMoEExpertsMonolithic``` doesn't.
Based on this we can avoid add the CPU specific check. | 4bc83323f2ea8e85c87ae5fb5ff2d792a8f61f9d | 96535b255c5e430c7faa3213b6038ccf79743f91 | diff --git a/vllm/model_executor/layers/quantization/auto_gptq.py b/vllm/model_executor/layers/quantization/auto_gptq.py
index 1821fd5c7f71..459a61583271 100644
--- a/vllm/model_executor/layers/quantization/auto_gptq.py
+++ b/vllm/model_executor/layers/quantization/auto_gptq.py
@@ -16,6 +16,7 @@
)
from vllm.model_exe... | [
"vllm/model_executor/layers/quantization/auto_gptq.py"
] | [
{
"comment": "In current MoE backend abstraction, ```FusedMoEExpertsModular``` requires workspace allocation and ```FusedMoEExpertsMonolithic``` doesn't.\n\nBased on this we can avoid add the CPU specific check.",
"path": "vllm/model_executor/layers/quantization/auto_gptq.py",
"hunk": "@@ -640,8 +640,13... | true | ||
vllm-project/vllm | 43,409 | comment_to_fix | [CPU] Support CPU W4A16 INT4 MoE | Better to check ```apply_router_weight_on_input``` here to avoid silent accuracy bug.
In most models it is ```False```, only ```Llama4``` uses it. | 4bc83323f2ea8e85c87ae5fb5ff2d792a8f61f9d | 96535b255c5e430c7faa3213b6038ccf79743f91 | diff --git a/vllm/model_executor/layers/fused_moe/experts/cpu_moe.py b/vllm/model_executor/layers/fused_moe/experts/cpu_moe.py
index 84740fc05703..11ed775f28e5 100644
--- a/vllm/model_executor/layers/fused_moe/experts/cpu_moe.py
+++ b/vllm/model_executor/layers/fused_moe/experts/cpu_moe.py
@@ -5,7 +5,12 @@
import torc... | [
"vllm/model_executor/layers/fused_moe/experts/cpu_moe.py"
] | [
{
"comment": "Better to check ```apply_router_weight_on_input``` here to avoid silent accuracy bug.\nIn most models it is ```False```, only ```Llama4``` uses it.",
"path": "vllm/model_executor/layers/fused_moe/experts/cpu_moe.py",
"hunk": "@@ -318,3 +324,201 @@ def apply(\n limit,\n ... | true | ||
vllm-project/vllm | 43,409 | comment_to_fix | [CPU] Support CPU W4A16 INT4 MoE | Hmmm, I'm not sure whether this replacement is safe because different backends may have different behaviors. I would suggest to keep it as original. | 4bc83323f2ea8e85c87ae5fb5ff2d792a8f61f9d | 96535b255c5e430c7faa3213b6038ccf79743f91 | diff --git a/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe/compressed_tensors_moe_wna16_marlin.py b/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe/compressed_tensors_moe_wna16_marlin.py
index 2a98d444afd5..a69d2a594ad6 100644
--- a/vllm/model_execut... | [
"vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe/compressed_tensors_moe_wna16_marlin.py"
] | [
{
"comment": "Hmmm, I'm not sure whether this replacement is safe because different backends may have different behaviors. I would suggest to keep it as original. ",
"path": "vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe/compressed_tensors_moe_wna16_marlin.py",
"hunk"... | true | ||
vllm-project/vllm | 43,409 | comment_to_fix | [CPU] Support CPU W4A16 INT4 MoE | New arguments here, need to update ```CPUExpertsInt4``` | 4bc83323f2ea8e85c87ae5fb5ff2d792a8f61f9d | 96535b255c5e430c7faa3213b6038ccf79743f91 | diff --git a/vllm/model_executor/layers/fused_moe/oracle/int_wna16.py b/vllm/model_executor/layers/fused_moe/oracle/int_wna16.py
index 6ad60d62e97d..8de6269e2e9e 100644
--- a/vllm/model_executor/layers/fused_moe/oracle/int_wna16.py
+++ b/vllm/model_executor/layers/fused_moe/oracle/int_wna16.py
@@ -45,6 +45,7 @@
class ... | [
"vllm/model_executor/layers/fused_moe/oracle/int_wna16.py"
] | [
{
"comment": "New arguments here, need to update ```CPUExpertsInt4```",
"path": "vllm/model_executor/layers/fused_moe/oracle/int_wna16.py",
"hunk": "",
"resolving_sha": "96535b255c5e430c7faa3213b6038ccf79743f91",
"resolving_diff": "diff --git a/vllm/model_executor/layers/fused_moe/oracle/int_wna... | true | ||
vllm-project/vllm | 45,491 | issue_to_patch | Treat null completion max_tokens like the default | This fixes a Schemathesis failure in the OpenAI API server job:
```text
AMD: Entrypoints Integration (API Server OpenAI - Part 1) (mi325_1)
tests/entrypoints/openai/test_openai_schema.py::test_openapi_stateless[POST /v1/completions]
```
Buildkite:
https://buildkite.com/vllm/ci/builds/71260/canvas?jid=019eb35... | 17ee5b1ac5dd61fa89bc4321ef54b0a790a45db3 | 574114c58c74222efc52da5edc5ff1608a6fedf7 | diff --git a/vllm/entrypoints/openai/completion/protocol.py b/vllm/entrypoints/openai/completion/protocol.py
index 30a4f20084e1..1d61ca3c5980 100644
--- a/vllm/entrypoints/openai/completion/protocol.py
+++ b/vllm/entrypoints/openai/completion/protocol.py
@@ -346,6 +346,14 @@ def to_sampling_params(
thinkin... | [
"vllm/entrypoints/openai/completion/protocol.py"
] | [] | true | ||
vllm-project/vllm | 44,774 | issue_to_patch | [KV Connector] Mooncake store: prefix-cache retention interval for sparse attention | Follow-up of #43447.
## Summary
Adds prefix-cache retention support to the Mooncake KV store's `store_mask`, so the store writes exactly the blocks a future local prefix-cache hit can actually consume — and no more.
Previously `store_mask` computed a per-lcm-region template (via a dummy-hash pass through `find_longe... | 82d6b59f0411b70ad2b5a74a24f808270e0af588 | f3e9a081d341b5243f7a954dd27e505eb0002b28 | diff --git a/vllm/distributed/kv_transfer/kv_connector/v1/mooncake/store/coordinator.py b/vllm/distributed/kv_transfer/kv_connector/v1/mooncake/store/coordinator.py
index ad5281409664..227575c92677 100644
--- a/vllm/distributed/kv_transfer/kv_connector/v1/mooncake/store/coordinator.py
+++ b/vllm/distributed/kv_transfer... | [
"tests/v1/kv_connector/unit/test_mooncake_store_coordinator.py",
"vllm/distributed/kv_transfer/kv_connector/v1/mooncake/store/coordinator.py",
"vllm/distributed/kv_transfer/kv_connector/v1/mooncake/store/data.py",
"vllm/distributed/kv_transfer/kv_connector/v1/mooncake/store/worker.py"
] | [] | diff --git a/tests/v1/kv_connector/unit/test_mooncake_store_coordinator.py b/tests/v1/kv_connector/unit/test_mooncake_store_coordinator.py
index 492a905ed16c..677e4de22b2a 100644
--- a/tests/v1/kv_connector/unit/test_mooncake_store_coordinator.py
+++ b/tests/v1/kv_connector/unit/test_mooncake_store_coordinator.py
@@ -1... | true | |
vllm-project/vllm | 44,103 | issue_to_patch | [Bugfix][Mooncake] Fix per-group block_size/block_hash and group_idx in MooncakeStoreConnector KV events | ## Purpose
`MooncakeStoreConnector` emits `BlockStored` KV events so external subscribers can track which KV blocks are present in the store. In hybrid / multi-group (HMA) configurations — and whenever a group's `block_size` is larger than the connector's `hash_block_size` — three fields of the emitted event were wron... | 32f34d393524d18c7ec9c4add0be7731cde372dc | 8114c9c4412759cdb216ff8d1c5edb752e333c66 | diff --git a/vllm/distributed/kv_transfer/kv_connector/v1/mooncake/store/data.py b/vllm/distributed/kv_transfer/kv_connector/v1/mooncake/store/data.py
index 2a625c062779..b26e6835a9c9 100644
--- a/vllm/distributed/kv_transfer/kv_connector/v1/mooncake/store/data.py
+++ b/vllm/distributed/kv_transfer/kv_connector/v1/moon... | [
"tests/v1/kv_connector/unit/test_mooncake_store_scheduler.py",
"tests/v1/kv_connector/unit/test_mooncake_store_worker.py",
"vllm/distributed/kv_transfer/kv_connector/v1/mooncake/store/data.py",
"vllm/distributed/kv_transfer/kv_connector/v1/mooncake/store/scheduler.py",
"vllm/distributed/kv_transfer/kv_conne... | [] | diff --git a/tests/v1/kv_connector/unit/test_mooncake_store_scheduler.py b/tests/v1/kv_connector/unit/test_mooncake_store_scheduler.py
index 5ee4620d5ace..ac36005c63ef 100644
--- a/tests/v1/kv_connector/unit/test_mooncake_store_scheduler.py
+++ b/tests/v1/kv_connector/unit/test_mooncake_store_scheduler.py
@@ -16,7 +16,... | true | |
vllm-project/vllm | 44,165 | issue_to_patch | [Core][Refactor]: thread `scheduler_block_size` into KVCacheManager and KVCacheCoordinator | ## Purpose
This is a small, behavior-preserving refactor that threads an explicit `scheduler_block_size` through `KVCacheManager` → `KVCacheCoordinator` → `SingleTypeKVCacheManager`, instead of having `HybridKVCacheCoordinator` recompute the LCM of group block sizes internally.
Today the scheduler already resolves th... | 8c3cc98cffd31b910c41b11076e8c175fc6dabe9 | 8c8a6e5c90056107e73042416f235a5964b36bc1 | diff --git a/vllm/v1/core/kv_cache_coordinator.py b/vllm/v1/core/kv_cache_coordinator.py
index c5e8953745ad..0336ecdfd283 100644
--- a/vllm/v1/core/kv_cache_coordinator.py
+++ b/vllm/v1/core/kv_cache_coordinator.py
@@ -2,7 +2,6 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from abc import AB... | [
"tests/v1/core/test_kv_cache_utils.py",
"tests/v1/core/test_prefix_caching.py",
"tests/v1/core/test_single_type_kv_cache_manager.py",
"tests/v1/kv_connector/unit/offloading_connector/utils.py",
"vllm/v1/core/kv_cache_coordinator.py",
"vllm/v1/core/kv_cache_manager.py",
"vllm/v1/core/sched/scheduler.py",... | [
{
"comment": "Could/should we add an assert here for this?",
"path": "vllm/v1/core/kv_cache_coordinator.py",
"hunk": "@@ -40,12 +39,16 @@ def __init__(\n enable_kv_cache_events: bool,\n dcp_world_size: int,\n pcp_world_size: int,\n+ scheduler_block_size: int,\n has... | diff --git a/tests/v1/core/test_kv_cache_utils.py b/tests/v1/core/test_kv_cache_utils.py
index 94e9f6f4c100..68ad7bc42ef0 100644
--- a/tests/v1/core/test_kv_cache_utils.py
+++ b/tests/v1/core/test_kv_cache_utils.py
@@ -1447,7 +1447,10 @@ def test_allocate_with_lookahead():
# Test case 1: Requires additional look... | true | |
vllm-project/vllm | 42,296 | issue_to_patch | [Feat][KVConnector] Support DSV4 in SimpleCPUOffloadBackend | ## Purpose
Based on https://github.com/vllm-project/vllm/pull/42258, need to merge #42258 first.
Resolving `scheduler_block_size` and `hash_block_size` in SimpleCPUOffloadBackend, and skipping impossible SWA blocks from prefix caching. The fix mainly applies to the eager mode, as lazy mode works naturally.
## ... | d0a100c87af832ad97ade60b8ec7610018a08427 | fee02833e901ca2da94584c4997caa23ea364973 | diff --git a/vllm/distributed/kv_transfer/kv_connector/v1/simple_cpu_offload_connector.py b/vllm/distributed/kv_transfer/kv_connector/v1/simple_cpu_offload_connector.py
index be8ee8e4e95a..15904da9e531 100644
--- a/vllm/distributed/kv_transfer/kv_connector/v1/simple_cpu_offload_connector.py
+++ b/vllm/distributed/kv_tr... | [
"tests/v1/simple_kv_offload/test_scheduler.py",
"vllm/distributed/kv_transfer/kv_connector/v1/simple_cpu_offload_connector.py",
"vllm/v1/simple_kv_offload/manager.py"
] | [] | diff --git a/tests/v1/simple_kv_offload/test_scheduler.py b/tests/v1/simple_kv_offload/test_scheduler.py
index 104b0184dbe2..970e16e52798 100644
--- a/tests/v1/simple_kv_offload/test_scheduler.py
+++ b/tests/v1/simple_kv_offload/test_scheduler.py
@@ -153,6 +153,8 @@ def make_scheduler(
vllm_config=vllm_config,... | true | |
vllm-project/vllm | 42,258 | issue_to_patch | [Core][DSV4] Skip caching SWA blocks that can never serve a prefix-cache hit | ## Purpose
DeepSeek-V4 pairs full-attention layers with SWA layers with different block sizes and window sizes. The full attn layers have block size of 256, while SWA layers and compressors have block sizes of 64, 4, or 8.
- Within each 256-aligned segment, only the trailing `tail = ceil((sliding_window - 1) / bl... | 4cfcc0866f233c0f223b78eaa205f936d5522324 | c106593ea195170823825a5d6ab666bf16e0e40b | diff --git a/vllm/v1/core/block_pool.py b/vllm/v1/core/block_pool.py
index 9097079ef33a..513e4bf380b9 100644
--- a/vllm/v1/core/block_pool.py
+++ b/vllm/v1/core/block_pool.py
@@ -216,6 +216,7 @@ def cache_full_blocks(
num_full_blocks: int,
block_size: int,
kv_cache_group_id: int,
+ blo... | [
"tests/v1/core/test_prefix_caching.py",
"vllm/v1/core/block_pool.py",
"vllm/v1/core/kv_cache_coordinator.py",
"vllm/v1/core/single_type_kv_cache_manager.py"
] | [
{
"comment": "\n\nFor hybrid models with shared physical blocks (like DeepSeek-V4), multiple KV cache groups will attempt to cache the same physical block. The current logic will trigger an assertion failure at line 273 because `blk.block_hash` w... | diff --git a/tests/v1/core/test_prefix_caching.py b/tests/v1/core/test_prefix_caching.py
index 0a58fa2f4fa9..546412b1d2f8 100644
--- a/tests/v1/core/test_prefix_caching.py
+++ b/tests/v1/core/test_prefix_caching.py
@@ -2528,6 +2528,145 @@ def test_different_block_size():
assert num_computed_tokens == 4 * 16
+d... | true | |
vllm-project/vllm | 42,258 | comment_to_fix | [Core][DSV4] Skip caching SWA blocks that can never serve a prefix-cache hit | 
For hybrid models with shared physical blocks (like DeepSeek-V4), multiple KV cache groups will attempt to cache the same physical block. The current logic will trigger an assertion failure at line 273 because `blk.block_hash` will already be set by th... | 4cfcc0866f233c0f223b78eaa205f936d5522324 | c106593ea195170823825a5d6ab666bf16e0e40b | diff --git a/vllm/v1/core/block_pool.py b/vllm/v1/core/block_pool.py
index 9097079ef33a..513e4bf380b9 100644
--- a/vllm/v1/core/block_pool.py
+++ b/vllm/v1/core/block_pool.py
@@ -216,6 +216,7 @@ def cache_full_blocks(
num_full_blocks: int,
block_size: int,
kv_cache_group_id: int,
+ blo... | [
"vllm/v1/core/block_pool.py"
] | [
{
"comment": "\n\nFor hybrid models with shared physical blocks (like DeepSeek-V4), multiple KV cache groups will attempt to cache the same physical block. The current logic will trigger an assertion failure at line 273 because `blk.block_hash` w... | true | ||
vllm-project/vllm | 42,258 | comment_to_fix | [Core][DSV4] Skip caching SWA blocks that can never serve a prefix-cache hit | 
The new tests do not cover the case of hybrid models with shared physical blocks (like DeepSeek-V4). In such models, multiple groups will attempt to cache the same block, which will trigger an assertion failure in `BlockPool.cache_full_blocks` as curre... | 4cfcc0866f233c0f223b78eaa205f936d5522324 | c106593ea195170823825a5d6ab666bf16e0e40b | diff --git a/tests/v1/core/test_prefix_caching.py b/tests/v1/core/test_prefix_caching.py
index 0a58fa2f4fa9..546412b1d2f8 100644
--- a/tests/v1/core/test_prefix_caching.py
+++ b/tests/v1/core/test_prefix_caching.py
@@ -2528,6 +2528,145 @@ def test_different_block_size():
assert num_computed_tokens == 4 * 16
+d... | [
"tests/v1/core/test_prefix_caching.py"
] | [
{
"comment": "\n\nThe new tests do not cover the case of hybrid models with shared physical blocks (like DeepSeek-V4). In such models, multiple groups will attempt to cache the same block, which will trigger an assertion failure in `BlockPool.cac... | true | ||
vllm-project/vllm | 45,460 | issue_to_patch | [Bug]: thread-pool tokenizer (maybe_make_thread_pool) unpickles to None — breaks get_tokenizer() across Ray / multiprocessing
### Your current environment
<details>
<summary>The output of <code>python collect_env.py</code>(trimmed to key entries)</summary>
```text
OS : Ubuntu 24.04.4 LTS (x86_64)
Pyth... | [Bugfix] Return the tokenizer from maybe_make_thread_pool so it survives pickling | ## Purpose
Fix #45433: tokenizers wrapped by `maybe_make_thread_pool` (every fast tokenizer returned by `LLM.get_tokenizer()` / `AsyncLLM.get_tokenizer()` since the thread-pool wrapper landed) **unpickle to `None`**, breaking tokenizer retrieval across Ray actors, multiprocessing, and any cloudpickle path.
Mechanism:... | ff5a30cfac59c9c753b6340a59c6d8ed668752f0 | 6018a7ec83baa9b6d1a7aaa27e71b5aea90e045f | diff --git a/vllm/tokenizers/hf.py b/vllm/tokenizers/hf.py
index b4248e229a68..45370bbb394c 100644
--- a/vllm/tokenizers/hf.py
+++ b/vllm/tokenizers/hf.py
@@ -99,6 +99,9 @@ def __reduce__(self):
TokenizerPool.__name__ = f"TokenizerPool{og_tokenizer.__class__.__name__}"
tokenizer.__class__ = TokenizerPool
+ ... | [
"tests/tokenizers_/test_hf.py",
"vllm/tokenizers/hf.py"
] | [
{
"comment": "Is this pooled_tokenizer with the tokenizer_pool closure really picklable? \n\nPerhaps it is, (Because the thread pool is not inside this closure, which is not picklable.)\nbut it's really not a good design. \n\nOur maybe_make_thread_pool has too much wrapping(magic), making it unusable for downst... | diff --git a/tests/tokenizers_/test_hf.py b/tests/tokenizers_/test_hf.py
index c1238900ce0d..3ccbbd73e7a6 100644
--- a/tests/tokenizers_/test_hf.py
+++ b/tests/tokenizers_/test_hf.py
@@ -7,7 +7,11 @@
from transformers import AutoTokenizer
from vllm.tokenizers import TokenizerLike
-from vllm.tokenizers.hf import get... | true |
vllm-project/vllm | 44,572 | issue_to_patch | [Perf] SM90 cutlass fp8 mm supports odd M by swap_ab, 180~290% kernel performance improvement | ## Purpose
A follow up for https://github.com/vllm-project/vllm/pull/43706, we now directly implement in kernel level.
The swap_ab logic is from SM100/SM120 currently
## Test
Acc covered in unit tests
Perf:
```py
import time
import statistics
import torch
from vllm import _custom_ops as ops
fro... | a30addc7548a9a8b9b3323a7bc3eb7d7c4895d1c | 16022a2d37a7be3a8db37ff53e586a2ec2aec560 | diff --git a/csrc/libtorch_stable/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm90_fp8_dispatch.cuh b/csrc/libtorch_stable/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm90_fp8_dispatch.cuh
index cf62e81fd75b..529b28ceece5 100644
--- a/csrc/libtorch_stable/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm... | [
"csrc/libtorch_stable/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm90_fp8_dispatch.cuh",
"tests/kernels/quantization/test_cutlass_scaled_mm.py",
"vllm/model_executor/kernels/linear/scaled_mm/cutlass.py"
] | [] | diff --git a/tests/kernels/quantization/test_cutlass_scaled_mm.py b/tests/kernels/quantization/test_cutlass_scaled_mm.py
index a937c30fed74..25893311afca 100644
--- a/tests/kernels/quantization/test_cutlass_scaled_mm.py
+++ b/tests/kernels/quantization/test_cutlass_scaled_mm.py
@@ -245,8 +245,6 @@ def test_cutlass_fp8_... | true | |
vllm-project/vllm | 45,468 | issue_to_patch | [Bug]: DiffusionGemma structured JSON outputs fail with xgrammar FSM rejection
### Your current environment
Docker image:
```text
vllm/vllm-openai:pr45163-a275a28e-dgemma-fullpy
```
vLLM version from server log:
```text
v0.22.1rc1.dev332+g2c9c07c85
VLLM_BUILD_COMMIT=2c9c07c85e56c799afffd5a671a8a0bace377a39
```
Ha... | [Bugfix] Reject structured outputs for diffusion decoders with a clear error | ## Purpose
Fixes #45436.
`nvidia/diffusiongemma-26B-A4B-it-NVFP4` (DiffusionGemma, added in #45163) fails with an
HTTP 500 when a structured output constraint is supplied (`response_format: json_schema`
or top-level `structured_outputs`): the request starts generating and is then terminated
mid-flight by an xgrammar ... | e3e31e54b05391d21a4b492d3bde612f47696975 | 148adb75948e5e741efefa53ec048ee959f481c3 | diff --git a/vllm/sampling_params.py b/vllm/sampling_params.py
index 17204093ab17..2786ca8c5c19 100644
--- a/vllm/sampling_params.py
+++ b/vllm/sampling_params.py
@@ -720,7 +720,9 @@ def verify(
self._validate_logits_processors(model_config)
self._validate_allowed_token_ids(tokenizer)
self._v... | [
"tests/v1/structured_output/test_validation.py",
"vllm/sampling_params.py"
] | [] | diff --git a/tests/v1/structured_output/test_validation.py b/tests/v1/structured_output/test_validation.py
new file mode 100644
index 000000000000..1b8581c1c621
--- /dev/null
+++ b/tests/v1/structured_output/test_validation.py
@@ -0,0 +1,50 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright ... | true |
vllm-project/vllm | 45,536 | issue_to_patch | Fix docs build on `main` | We will have to do this periodically now that we've stopped building docs on every PR | 470229c37efaf69c86e8bc97482b0b1ff7551c65 | d98a0aa536482056bb7739fce17cbb409adb328d | diff --git a/vllm/model_executor/layers/fused_moe/experts/cpu_moe.py b/vllm/model_executor/layers/fused_moe/experts/cpu_moe.py
index 11ed775f28e5..cd67207b7100 100644
--- a/vllm/model_executor/layers/fused_moe/experts/cpu_moe.py
+++ b/vllm/model_executor/layers/fused_moe/experts/cpu_moe.py
@@ -351,8 +351,7 @@ def prepa... | [
"vllm/model_executor/layers/fused_moe/experts/cpu_moe.py",
"vllm/model_executor/layers/fusion/__init__.py"
] | [] | true | ||
vllm-project/vllm | 45,478 | issue_to_patch | [CI Bug] Fix `ValueError: There is no module or parameter named 'model.vision_tower.vision_model'` | ## Purpose
Fixes https://buildkite.com/vllm/ci/builds/71235#019ebd0b-5a0c-44e8-9f0c-7a7965833cc8
```bash
(EngineCore pid=3055546) File "/home/yewentao256/vllm-source/vllm/tracing/otel.py", line 178, in sync_wrapper
(EngineCore pid=3055546) return func(*args, **kwargs)
(EngineCore pid=3055546) ... | 1a369783e9a09cfd9ebed9799a7b8bbffdc9896f | 9aff242289c8b42a07495d8a5dea3ece7026c02c | diff --git a/vllm/model_executor/models/transformers/base.py b/vllm/model_executor/models/transformers/base.py
index 234ae9570b21..55d946004974 100644
--- a/vllm/model_executor/models/transformers/base.py
+++ b/vllm/model_executor/models/transformers/base.py
@@ -303,17 +303,13 @@ def _create_hf_to_vllm_mapper(self):
... | [
"vllm/model_executor/models/transformers/base.py",
"vllm/model_executor/models/utils.py"
] | [] | true | ||
vllm-project/vllm | 45,003 | issue_to_patch | [Frontend] Support strict mode for tool calling | Co-authored-by: @cjackal <44624812+cjackal@users.noreply.github.com>
## Purpose
[Frontend] Support strict mode for tool calling
## Test Plan
I tested it locally with Minimax 2.5,Qwen2.5, Qwen3.5, Qwen3.6, Qwen3, and DeepSeek V3.2.
I also tested the tool_choice modes required, auto, and named tool selecti... | 39dee1114a2cd183a9fb72b561808b385b6c9daa | 4d50f7de4c38a9ebacfac8956076d2996a6dfb71 | diff --git a/docs/features/tool_calling.md b/docs/features/tool_calling.md
index d1a56e83cd4c..43010c406f53 100644
--- a/docs/features/tool_calling.md
+++ b/docs/features/tool_calling.md
@@ -115,18 +115,28 @@ Whether vLLM enforces the tool parameter schema during generation depends on the
| --- | --- | --- |
| Named ... | [
"docs/features/tool_calling.md",
"requirements/common.txt",
"requirements/test/rocm.txt",
"tests/entrypoints/openai/chat_completion/test_completion_with_function_calling.py",
"tests/entrypoints/openai/responses/conftest.py",
"tests/parser/test_parse.py",
"tests/tool_parsers/test_qwen3coder_tool_parser.p... | [
{
"comment": "Do we want to disallow non-stag usage? We have some FSM advancement failure type blockers (which emits 500 status error on client side) for (xgrammar + PP) & (xgrammar + speculative decoding), so it may hurt UX if we have no way to disable strict tool calling feature.",
"path": "vllm/envs.py",... | diff --git a/requirements/test/rocm.txt b/requirements/test/rocm.txt
index ce18ce456cc2..a6fc7242174e 100644
--- a/requirements/test/rocm.txt
+++ b/requirements/test/rocm.txt
@@ -1367,7 +1367,7 @@ word2number==1.1
# via lm-eval
wrapt==2.1.2
# via smart-open
-xgrammar==0.2.0
+xgrammar==0.2.1
# via
#... | true | |
vllm-project/vllm | 45,003 | comment_to_fix | [Frontend] Support strict mode for tool calling | Do we want to disallow non-stag usage? We have some FSM advancement failure type blockers (which emits 500 status error on client side) for (xgrammar + PP) & (xgrammar + speculative decoding), so it may hurt UX if we have no way to disable strict tool calling feature. | 39dee1114a2cd183a9fb72b561808b385b6c9daa | 4d50f7de4c38a9ebacfac8956076d2996a6dfb71 | diff --git a/vllm/envs.py b/vllm/envs.py
index 479aab2323c3..dfebcd27ae82 100755
--- a/vllm/envs.py
+++ b/vllm/envs.py
@@ -200,6 +200,7 @@
MOONCAKE_REQUESTER_LOCAL_HOSTNAME: str | None = None
VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: int = 163840
VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS: int = 1
+ VLLM_ENFORCE... | [
"vllm/envs.py"
] | [
{
"comment": "Do we want to disallow non-stag usage? We have some FSM advancement failure type blockers (which emits 500 status error on client side) for (xgrammar + PP) & (xgrammar + speculative decoding), so it may hurt UX if we have no way to disable strict tool calling feature.",
"path": "vllm/envs.py",... | true | ||
vllm-project/vllm | 45,003 | comment_to_fix | [Frontend] Support strict mode for tool calling | We'd want the whole grammar with the reasoning part, but at the level of the tool parser, we don't have this knowledge, which is why I think the tag logic would be better to reside in https://github.com/vllm-project/vllm/blob/main/vllm/parser/abstract_parser.py#L616, so that we can check reasoning kwargs, etc. Wdyt? | 39dee1114a2cd183a9fb72b561808b385b6c9daa | 4d50f7de4c38a9ebacfac8956076d2996a6dfb71 | diff --git a/vllm/tool_parsers/abstract_tool_parser.py b/vllm/tool_parsers/abstract_tool_parser.py
index 94543b82350b..c2face916808 100644
--- a/vllm/tool_parsers/abstract_tool_parser.py
+++ b/vllm/tool_parsers/abstract_tool_parser.py
@@ -6,6 +6,7 @@
import os
from collections.abc import Callable, Sequence
from func... | [
"vllm/tool_parsers/abstract_tool_parser.py"
] | [
{
"comment": "We'd want the whole grammar with the reasoning part, but at the level of the tool parser, we don't have this knowledge, which is why I think the tag logic would be better to reside in https://github.com/vllm-project/vllm/blob/main/vllm/parser/abstract_parser.py#L616, so that we can check reasoning... | true | ||
vllm-project/vllm | 45,003 | comment_to_fix | [Frontend] Support strict mode for tool calling | Do you want to leave the response api part as follow up? | 39dee1114a2cd183a9fb72b561808b385b6c9daa | 4d50f7de4c38a9ebacfac8956076d2996a6dfb71 | diff --git a/vllm/parser/abstract_parser.py b/vllm/parser/abstract_parser.py
index 4fe7b7ec4d56..474dec5bd139 100644
--- a/vllm/parser/abstract_parser.py
+++ b/vllm/parser/abstract_parser.py
@@ -25,6 +25,7 @@
from vllm.logger import init_logger
from vllm.parser.metrics import record_tool_parser_invocation
from vllm.... | [
"vllm/parser/abstract_parser.py"
] | [
{
"comment": "Do you want to leave the response api part as follow up?",
"path": "vllm/parser/abstract_parser.py",
"hunk": "@@ -618,10 +619,47 @@ def adjust_request(\n ) -> ChatCompletionRequest | ResponsesRequest:\n if self._reasoning_parser is not None:\n request = self._reason... | true | ||
vllm-project/vllm | 45,252 | issue_to_patch | [Security] Fix DoS via prompt_embeds on M-RoPE models | Replace the fatal assertion in _init_mrope_positions that crashes the EngineCore when prompt_token_ids is None. When only prompt_embeds are provided (e.g. /v1/completions with prompt=None + prompt_embeds), derive the sequence length from the embeddings tensor and use dummy token IDs.
This is safe because get_mrope_... | 272c16953eac7c46db7719d284d8a0ff19e63446 | 04b15c8899cb632721ab20a13bc3c3abbdcb8b93 | diff --git a/vllm/v1/worker/gpu_model_runner.py b/vllm/v1/worker/gpu_model_runner.py
index cb607c0b7b06..afda4ec0bb00 100644
--- a/vllm/v1/worker/gpu_model_runner.py
+++ b/vllm/v1/worker/gpu_model_runner.py
@@ -1588,9 +1588,6 @@ def _update_streaming_request(
def _init_mrope_positions(self, req_state: CachedReques... | [
"tests/v1/worker/test_mrope_prompt_embeds.py",
"vllm/v1/worker/gpu_model_runner.py"
] | [] | diff --git a/tests/v1/worker/test_mrope_prompt_embeds.py b/tests/v1/worker/test_mrope_prompt_embeds.py
new file mode 100644
index 000000000000..209b88f5222f
--- /dev/null
+++ b/tests/v1/worker/test_mrope_prompt_embeds.py
@@ -0,0 +1,79 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contri... | true | |
vllm-project/vllm | 45,118 | issue_to_patch | [Security] Add timeout guard for regex compilation in structured outp… | ## Summary
- Add `compile_regex_with_timeout()` utility that wraps regex compilation in a thread with a deadline, preventing ReDoS attacks where adversarial patterns cause exponential DFA state-space explosion.
- Apply the timeout to `compiler.compile_regex()` in the xgrammar backend, `oc.Index()` in the outlines b... | 272c16953eac7c46db7719d284d8a0ff19e63446 | 5c6ec2cc60d253d4d4341c0606fc3c0bbed9c930 | diff --git a/vllm/envs.py b/vllm/envs.py
index dfebcd27ae82..b8d45fc67b1a 100755
--- a/vllm/envs.py
+++ b/vllm/envs.py
@@ -188,6 +188,7 @@
VLLM_FLASHINFER_ALLREDUCE_BACKEND: Literal["auto", "trtllm", "mnnvl"] = "auto"
VLLM_FLASHINFER_WORKSPACE_BUFFER_SIZE: int = 394 * 1024 * 1024
VLLM_XGRAMMAR_CACHE_MB: ... | [
"tests/v1/structured_output/test_regex_compilation_timeout.py",
"vllm/envs.py",
"vllm/v1/structured_output/backend_outlines.py",
"vllm/v1/structured_output/backend_xgrammar.py",
"vllm/v1/structured_output/utils.py"
] | [
{
"comment": "🟠 **Severity: HIGH**\n\nUsing `ThreadPoolExecutor` as a context manager defeats the timeout: `__exit__` calls `shutdown(wait=True)`, which blocks until the spawned thread completes. An adversarial regex that triggers exponential DFA blowup will still hang the caller indefinitely despite `future.r... | diff --git a/tests/v1/structured_output/test_regex_compilation_timeout.py b/tests/v1/structured_output/test_regex_compilation_timeout.py
new file mode 100644
index 000000000000..b0eaeed95ee5
--- /dev/null
+++ b/tests/v1/structured_output/test_regex_compilation_timeout.py
@@ -0,0 +1,61 @@
+# SPDX-License-Identifier: Apa... | true | |
vllm-project/vllm | 45,118 | comment_to_fix | [Security] Add timeout guard for regex compilation in structured outp… | 🟠 **Severity: HIGH**
Using `ThreadPoolExecutor` as a context manager defeats the timeout: `__exit__` calls `shutdown(wait=True)`, which blocks until the spawned thread completes. An adversarial regex that triggers exponential DFA blowup will still hang the caller indefinitely despite `future.result(timeout=...)` firi... | 272c16953eac7c46db7719d284d8a0ff19e63446 | 5c6ec2cc60d253d4d4341c0606fc3c0bbed9c930 | diff --git a/vllm/v1/structured_output/utils.py b/vllm/v1/structured_output/utils.py
index f149ae845e31..d30dcf26170d 100644
--- a/vllm/v1/structured_output/utils.py
+++ b/vllm/v1/structured_output/utils.py
@@ -6,7 +6,9 @@
import importlib.metadata
import os
import tempfile
-from typing import TYPE_CHECKING
+from co... | [
"vllm/v1/structured_output/utils.py"
] | [
{
"comment": "🟠 **Severity: HIGH**\n\nUsing `ThreadPoolExecutor` as a context manager defeats the timeout: `__exit__` calls `shutdown(wait=True)`, which blocks until the spawned thread completes. An adversarial regex that triggers exponential DFA blowup will still hang the caller indefinitely despite `future.r... | true | ||
vllm-project/vllm | 45,118 | comment_to_fix | [Security] Add timeout guard for regex compilation in structured outp… | ```suggestion
def compile_regex_with_timeout(fn: Callable[[str], _T], pattern: str) -> _T:
```
So then call sites would do
```py
compile_regex_with_timeout(xgr.Grammar.from_regex, so_params.regex)
```
instead of
```py
compile_regex_with_timeout(xgr.Grammar.from_regex(so_params.regex), so_params.regex)
``` | 272c16953eac7c46db7719d284d8a0ff19e63446 | 5c6ec2cc60d253d4d4341c0606fc3c0bbed9c930 | diff --git a/vllm/v1/structured_output/utils.py b/vllm/v1/structured_output/utils.py
index f149ae845e31..d30dcf26170d 100644
--- a/vllm/v1/structured_output/utils.py
+++ b/vllm/v1/structured_output/utils.py
@@ -6,7 +6,9 @@
import importlib.metadata
import os
import tempfile
-from typing import TYPE_CHECKING
+from co... | [
"vllm/v1/structured_output/utils.py"
] | [
{
"comment": "```suggestion\ndef compile_regex_with_timeout(fn: Callable[[str], _T], pattern: str) -> _T:\n```\n\nSo then call sites would do\n\n```py\ncompile_regex_with_timeout(xgr.Grammar.from_regex, so_params.regex)\n```\n\ninstead of \n\n```py\ncompile_regex_with_timeout(xgr.Grammar.from_regex(so_params.re... | true | ||
vllm-project/vllm | 45,118 | comment_to_fix | [Security] Add timeout guard for regex compilation in structured outp… | Was this meant to be added? `compile_regex_with_timeout` will already raise if it needs to
```suggestion
``` | 272c16953eac7c46db7719d284d8a0ff19e63446 | 5c6ec2cc60d253d4d4341c0606fc3c0bbed9c930 | diff --git a/vllm/v1/structured_output/backend_xgrammar.py b/vllm/v1/structured_output/backend_xgrammar.py
index a92be3d44320..4f199a1a2735 100644
--- a/vllm/v1/structured_output/backend_xgrammar.py
+++ b/vllm/v1/structured_output/backend_xgrammar.py
@@ -19,6 +19,7 @@
)
from vllm.v1.structured_output.utils import (
... | [
"vllm/v1/structured_output/backend_xgrammar.py"
] | [
{
"comment": "Was this meant to be added? `compile_regex_with_timeout` will already raise if it needs to\n\n\n```suggestion\n```",
"path": "vllm/v1/structured_output/backend_xgrammar.py",
"hunk": "@@ -277,7 +281,12 @@ def validate_xgrammar_grammar(sampling_params: SamplingParams) -> None:\n \n if so... | true | ||
vllm-project/vllm | 45,118 | comment_to_fix | [Security] Add timeout guard for regex compilation in structured outp… | This test doesn't test what it says it does
```suggestion
``` | 272c16953eac7c46db7719d284d8a0ff19e63446 | 5c6ec2cc60d253d4d4341c0606fc3c0bbed9c930 | diff --git a/tests/v1/structured_output/test_regex_compilation_timeout.py b/tests/v1/structured_output/test_regex_compilation_timeout.py
new file mode 100644
index 000000000000..b0eaeed95ee5
--- /dev/null
+++ b/tests/v1/structured_output/test_regex_compilation_timeout.py
@@ -0,0 +1,61 @@
+# SPDX-License-Identifier: Apa... | [
"tests/v1/structured_output/test_regex_compilation_timeout.py"
] | [
{
"comment": "This test doesn't test what it says it does\n```suggestion\n```",
"path": "tests/v1/structured_output/test_regex_compilation_timeout.py",
"hunk": "@@ -0,0 +1,73 @@\n+# SPDX-License-Identifier: Apache-2.0\n+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project\n+\"\"\"Tests f... | true | ||
vllm-project/vllm | 45,467 | issue_to_patch | [Model Runner V2] Fix `openai.InternalServerError: Error code: 500 - 'list index out of range'` | ## Purpose
`VLLM_USE_V2_MODEL_RUNNER=1 pytest tests/entrypoints/openai/chat_completion/test_chat_echo.py::test_top_logprobs -xvs`
Will raise error
```bash
> completion = await client.chat.completions.create(
model=MODEL_NAME,
messages=messages,
max_tokens=1,
... | e3e31e54b05391d21a4b492d3bde612f47696975 | 61beacdf2099dc86c897adeaa3e1fade0c3a1226 | diff --git a/vllm/v1/worker/gpu/sample/states.py b/vllm/v1/worker/gpu/sample/states.py
index bf2f1ce78feb..fe4dee6a6b10 100644
--- a/vllm/v1/worker/gpu/sample/states.py
+++ b/vllm/v1/worker/gpu/sample/states.py
@@ -56,6 +56,8 @@ def add_request(self, req_idx: int, sampling_params: SamplingParams) -> None:
num_... | [
"vllm/v1/worker/gpu/sample/states.py"
] | [] | true | ||
vllm-project/vllm | 45,057 | issue_to_patch | [Bugfix] Handle HWC images in ImageProcessorItems.get_image_size | ## Purpose
`ImageProcessorItems.get_image_size()` unconditionally unpacks array shapes as CHW:
```python
_, h, w = image.shape
```
When the user passes an HWC image — the layout `np.array(PIL.Image)` produces, e.g. shape `(480, 640, 3)` — this returns `ImageSize(width=3, height=640)` instead of `ImageSize(width=640,... | 2c9c07c85e56c799afffd5a671a8a0bace377a39 | f66d69280330378e87433c49291b9f5e81aac648 | diff --git a/vllm/multimodal/parse.py b/vllm/multimodal/parse.py
index cdedd1942276..f4c72e060bff 100644
--- a/vllm/multimodal/parse.py
+++ b/vllm/multimodal/parse.py
@@ -334,7 +334,13 @@ def get_image_size(self, item_idx: int) -> ImageSize:
if isinstance(image, PILImage.Image):
return ImageSize(*... | [
"tests/multimodal/test_parse.py",
"vllm/multimodal/parse.py"
] | [] | diff --git a/tests/multimodal/test_parse.py b/tests/multimodal/test_parse.py
new file mode 100644
index 000000000000..6504cc6bcf3d
--- /dev/null
+++ b/tests/multimodal/test_parse.py
@@ -0,0 +1,51 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import nump... | true | |
vllm-project/vllm | 45,412 | issue_to_patch | [Doc] Fix uv dependency resolution failure for setuptools during CPU source builds (x86 & ARM) |
## Purpose
Fixes an immediate `uv` dependency resolution failure encountered when users follow the official documentation to build the CPU backend from source (impacting both Intel/AMD x86 and ARM AArch64).
The `uv pip install` commands utilize the `--torch-backend cpu` flag, pointing `uv` to the PyTorch whee... | 5b2943f5a6c5fb267d1b2029c666f9d6b0e4ebd6 | e82789159b1b522a1e0ea53fbd2a010508507c10 | diff --git a/docs/getting_started/installation/cpu.arm.inc.md b/docs/getting_started/installation/cpu.arm.inc.md
index f01ba429ee03..7a783b53c65f 100644
--- a/docs/getting_started/installation/cpu.arm.inc.md
+++ b/docs/getting_started/installation/cpu.arm.inc.md
@@ -96,8 +96,8 @@ cd vllm_source
Third, install required... | [
"docs/getting_started/installation/cpu.arm.inc.md",
"docs/getting_started/installation/cpu.x86.inc.md"
] | [] | true | ||
vllm-project/vllm | 43,606 | issue_to_patch | [Render] Add `/derender` endpoints for disaggregated postprocessing | ## Purpose
Add `POST /v1/chat/completions/derender` and `POST /v1/completions/derender` to the render server, completing the render to derender lifecycle for disaggregated serving (RFC #42729).
The generate server returns raw `token IDs` and `logprob` placeholders ("token_id:N" strings) because it has no tokenize... | c7aa3d263049ac9eefd0f59a10f5ecc6a78927df | e2dc40963ddc2020783470994b86faddf1425346 | diff --git a/vllm/entrypoints/openai/chat_completion/serving.py b/vllm/entrypoints/openai/chat_completion/serving.py
index 45b79c6a7ef0..b570b0c9871f 100644
--- a/vllm/entrypoints/openai/chat_completion/serving.py
+++ b/vllm/entrypoints/openai/chat_completion/serving.py
@@ -46,6 +46,7 @@
GenerationError,
Open... | [
"tests/entrypoints/serve/render/test_derender.py",
"vllm/entrypoints/openai/chat_completion/serving.py",
"vllm/entrypoints/openai/completion/serving.py",
"vllm/entrypoints/openai/engine/serving.py",
"vllm/entrypoints/serve/disagg/protocol.py",
"vllm/entrypoints/serve/render/api_router.py",
"vllm/entrypo... | [
{
"comment": "\n\nThe current implementation of `_resolve_token_placeholder` incorrectly resolves token strings and bytes for many tokenizers. `tokenizer.convert_ids_to_tokens` returns internal token representations (e.g., using ` ` for SentenceP... | diff --git a/tests/entrypoints/serve/render/test_derender.py b/tests/entrypoints/serve/render/test_derender.py
new file mode 100644
index 000000000000..a3006595c19a
--- /dev/null
+++ b/tests/entrypoints/serve/render/test_derender.py
@@ -0,0 +1,488 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Cop... | true | |
vllm-project/vllm | 43,606 | comment_to_fix | [Render] Add `/derender` endpoints for disaggregated postprocessing | 
The current implementation of `_resolve_token_placeholder` incorrectly resolves token strings and bytes for many tokenizers. `tokenizer.convert_ids_to_tokens` returns internal token representations (e.g., using ` ` for SentencePiece or `Ġ` for BPE), wh... | c7aa3d263049ac9eefd0f59a10f5ecc6a78927df | e2dc40963ddc2020783470994b86faddf1425346 | diff --git a/vllm/entrypoints/serve/render/serving.py b/vllm/entrypoints/serve/render/serving.py
index 6afb26d98435..05a291198337 100644
--- a/vllm/entrypoints/serve/render/serving.py
+++ b/vllm/entrypoints/serve/render/serving.py
@@ -1,5 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyrig... | [
"vllm/entrypoints/serve/render/serving.py"
] | [
{
"comment": "\n\nThe current implementation of `_resolve_token_placeholder` incorrectly resolves token strings and bytes for many tokenizers. `tokenizer.convert_ids_to_tokens` returns internal token representations (e.g., using ` ` for SentenceP... | true | ||
vllm-project/vllm | 43,606 | comment_to_fix | [Render] Add `/derender` endpoints for disaggregated postprocessing | Could we consolidate this code with the existing ones in serving chat/completions to avoid code drift? | c7aa3d263049ac9eefd0f59a10f5ecc6a78927df | e2dc40963ddc2020783470994b86faddf1425346 | diff --git a/vllm/entrypoints/serve/render/serving.py b/vllm/entrypoints/serve/render/serving.py
index 6afb26d98435..05a291198337 100644
--- a/vllm/entrypoints/serve/render/serving.py
+++ b/vllm/entrypoints/serve/render/serving.py
@@ -1,5 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyrig... | [
"vllm/entrypoints/serve/render/serving.py"
] | [
{
"comment": "Could we consolidate this code with the existing ones in serving chat/completions to avoid code drift?",
"path": "vllm/entrypoints/serve/render/serving.py",
"hunk": "@@ -438,6 +558,146 @@ def _make_request_with_harmony(\n \n return messages, [engine_input]\n \n+ async def derend... | true | ||
vllm-project/vllm | 43,606 | comment_to_fix | [Render] Add `/derender` endpoints for disaggregated postprocessing | Use `removeprefix` | c7aa3d263049ac9eefd0f59a10f5ecc6a78927df | e2dc40963ddc2020783470994b86faddf1425346 | diff --git a/vllm/entrypoints/openai/engine/serving.py b/vllm/entrypoints/openai/engine/serving.py
index f3e07336e824..5eb917ef96ae 100644
--- a/vllm/entrypoints/openai/engine/serving.py
+++ b/vllm/entrypoints/openai/engine/serving.py
@@ -452,7 +452,7 @@ def _get_decoded_token(
return_as_token_id: bool = False... | [
"vllm/entrypoints/openai/engine/serving.py"
] | [
{
"comment": "Use `removeprefix`",
"path": "vllm/entrypoints/openai/engine/serving.py",
"hunk": "@@ -597,6 +597,37 @@ def _is_model_supported(self, model_name: str | None) -> bool:\n return self.models.is_base_model(model_name)\n \n \n+def format_token_id_placeholder(token_id: int) -> str:\n+ ... | true | ||
vllm-project/vllm | 43,606 | comment_to_fix | [Render] Add `/derender` endpoints for disaggregated postprocessing | Is this just for debugging? I think we shouldn't log these normally to avoid overhead | c7aa3d263049ac9eefd0f59a10f5ecc6a78927df | e2dc40963ddc2020783470994b86faddf1425346 | diff --git a/vllm/entrypoints/serve/render/serving.py b/vllm/entrypoints/serve/render/serving.py
index 6afb26d98435..05a291198337 100644
--- a/vllm/entrypoints/serve/render/serving.py
+++ b/vllm/entrypoints/serve/render/serving.py
@@ -1,5 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyrig... | [
"vllm/entrypoints/serve/render/serving.py"
] | [
{
"comment": "Is this just for debugging? I think we shouldn't log these normally to avoid overhead",
"path": "vllm/entrypoints/serve/render/serving.py",
"hunk": "@@ -438,6 +540,146 @@ def _make_request_with_harmony(\n \n return messages, [engine_input]\n \n+ async def derender_chat_response(... | true | ||
vllm-project/vllm | 43,606 | comment_to_fix | [Render] Add `/derender` endpoints for disaggregated postprocessing | Use `warning_once` | c7aa3d263049ac9eefd0f59a10f5ecc6a78927df | e2dc40963ddc2020783470994b86faddf1425346 | diff --git a/vllm/entrypoints/openai/engine/serving.py b/vllm/entrypoints/openai/engine/serving.py
index f3e07336e824..5eb917ef96ae 100644
--- a/vllm/entrypoints/openai/engine/serving.py
+++ b/vllm/entrypoints/openai/engine/serving.py
@@ -452,7 +452,7 @@ def _get_decoded_token(
return_as_token_id: bool = False... | [
"vllm/entrypoints/openai/engine/serving.py"
] | [
{
"comment": "Use `warning_once`",
"path": "vllm/entrypoints/openai/engine/serving.py",
"hunk": "@@ -597,6 +597,38 @@ def _is_model_supported(self, model_name: str | None) -> bool:\n return self.models.is_base_model(model_name)\n \n \n+def format_token_id_placeholder(token_id: int) -> str:\n+ ... | true |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.