Merge branch 'main' into amd/agentx_dsv4_sgl_mtp_0717

Author: seungrokj

.github/workflows/README.md

@@ -157,6 +157,14 @@ they run all evals only. The primary label still controls canary/fail-fast.
 with both modifiers, are not. Default full sweeps, including default evals,
 are also reusable.
 
+## AgentX Fast Mode
+
+Add `agentx-fast` alongside one primary sweep label to run the 5-minute cache
+warmup and 20-minute profiling preset for single- and multi-node AgentX
+throughput jobs. Fixed-sequence throughput and eval jobs retain their canonical
+settings. Adding or removing the modifier restarts the active sweep. Fast-mode
+runs are not eligible for artifact reuse after merge.
+
 ## Reusing an Approved PR Full Sweep
 
 `[skip-sweep]` skips PR benchmark setup only; changelog and reuse checks still

.github/workflows/benchmark-multinode-tmpl.yml

@@ -274,8 +274,8 @@ jobs:
     timeout-minutes: 480
     name: >-
       p${{ inputs.priority }} | ${{ inputs.model-prefix }} ${{ inputs.precision }} ${{ inputs.runner }} ${{ inputs.framework == 'sglang' && 'sgl' || inputs.framework == 'dynamo-sglang' && 'dyn-sgl' || inputs.framework == 'sglang-disagg' && 'sgl-disagg' || inputs.framework }}
-      ${{ inputs.prefill-num-worker }}P (TP${{ inputs.prefill-tp }}${{ inputs.prefill-pp != '' && inputs.prefill-pp != '1' && format('/PP{0}', inputs.prefill-pp) || '' }}${{ inputs.prefill-dcp-size != '' && inputs.prefill-dcp-size != '1' && format('/DCP{0}', inputs.prefill-dcp-size) || '' }}${{ inputs.prefill-pcp-size != '' && inputs.prefill-pcp-size != '1' && format('/PCP{0}', inputs.prefill-pcp-size) || '' }}${{ inputs.prefill-ep != '' && inputs.prefill-ep != '1' && format('/EP{0}', inputs.prefill-ep) || '' }}${{ inputs.prefill-dp-attn == 'true' && '/DPA' || '' }})
-      x ${{ inputs.decode-num-worker }}D (TP${{ inputs.decode-tp }}${{ inputs.decode-pp != '' && inputs.decode-pp != '1' && format('/PP{0}', inputs.decode-pp) || '' }}${{ inputs.decode-dcp-size != '' && inputs.decode-dcp-size != '1' && format('/DCP{0}', inputs.decode-dcp-size) || '' }}${{ inputs.decode-pcp-size != '' && inputs.decode-pcp-size != '1' && format('/PCP{0}', inputs.decode-pcp-size) || '' }}${{ inputs.decode-ep != '' && inputs.decode-ep != '1' && format('/EP{0}', inputs.decode-ep) || '' }}${{ inputs.decode-dp-attn == 'true' && '/DPA' || '' }})
+      ${{ inputs.disagg == 'true' && format('{0}P ', inputs.prefill-num-worker) || '' }}(TP${{ inputs.prefill-tp }}${{ inputs.prefill-pp != '' && inputs.prefill-pp != '1' && format('/PP{0}', inputs.prefill-pp) || '' }}${{ inputs.prefill-dcp-size != '' && inputs.prefill-dcp-size != '1' && format('/DCP{0}', inputs.prefill-dcp-size) || '' }}${{ inputs.prefill-pcp-size != '' && inputs.prefill-pcp-size != '1' && format('/PCP{0}', inputs.prefill-pcp-size) || '' }}${{ inputs.prefill-ep != '' && inputs.prefill-ep != '1' && format('/EP{0}', inputs.prefill-ep) || '' }}${{ inputs.prefill-dp-attn == 'true' && '/DPA' || '' }})
+      ${{ inputs.disagg == 'true' && format('x {0}D (TP{1}{2}{3}{4}{5}{6})', inputs.decode-num-worker, inputs.decode-tp, inputs.decode-pp != '' && inputs.decode-pp != '1' && format('/PP{0}', inputs.decode-pp) || '', inputs.decode-dcp-size != '' && inputs.decode-dcp-size != '1' && format('/DCP{0}', inputs.decode-dcp-size) || '', inputs.decode-pcp-size != '' && inputs.decode-pcp-size != '1' && format('/PCP{0}', inputs.decode-pcp-size) || '', inputs.decode-ep != '' && inputs.decode-ep != '1' && format('/EP{0}', inputs.decode-ep) || '', inputs.decode-dp-attn == 'true' && '/DPA' || '') || '' }}
       ${{ inputs.spec-decoding != 'none' && inputs.spec-decoding || '' }}
       ${{ inputs.kv-offloading != '' && inputs.kv-offloading != 'none' && format('{0} KV offload', inputs.kv-offloading) || '' }}
       ${{ inputs.kv-offload-backend != '' && inputs.kv-offload-backend != 'none' && inputs.kv-offload-backend != 'default' && inputs.kv-offload-backend || '' }}

.github/workflows/run-sweep.yml

@@ -13,6 +13,7 @@ concurrency:
           github.event.label.name != 'full-sweep-fail-fast-no-canary' &&
           github.event.label.name != 'all-evals' &&
           github.event.label.name != 'evals-only' &&
+          github.event.label.name != 'agentx-fast' &&
           github.event.label.name != 'skip_queue' &&
           github.event.label.name != 'ci-patchwork' &&
           github.event.label.name != 'engine-patch' &&
@@ -55,6 +56,7 @@ jobs:
               github.event.label.name == 'full-sweep-fail-fast-no-canary' ||
               github.event.label.name == 'all-evals' ||
               github.event.label.name == 'evals-only' ||
+              github.event.label.name == 'agentx-fast' ||
               github.event.label.name == 'skip_queue' ||
               github.event.label.name == 'ci-patchwork' ||
               github.event.label.name == 'engine-patch' ||
@@ -136,6 +138,7 @@ jobs:
             github.event.action == 'synchronize' &&
             !github.event.pull_request.draft &&
             !contains(github.event.pull_request.labels.*.name, 'evals-only') &&
+            !contains(github.event.pull_request.labels.*.name, 'agentx-fast') &&
             (
               contains(github.event.pull_request.labels.*.name, 'full-sweep-enabled') ||
               contains(github.event.pull_request.labels.*.name, 'non-canary-full-sweep-enabled') ||
@@ -204,6 +207,7 @@ jobs:
                   github.event.label.name == 'full-sweep-fail-fast-no-canary' ||
                   github.event.label.name == 'all-evals' ||
                   github.event.label.name == 'evals-only' ||
+                  github.event.label.name == 'agentx-fast' ||
                   github.event.label.name == 'skip_queue' ||
                   github.event.label.name == 'ci-patchwork' ||
                   github.event.label.name == 'engine-patch' ||
@@ -622,6 +626,7 @@ jobs:
             disagg: ${{ 'false' }}
             run-eval: false
             scenario-type: agentic-coding
+            agentx-fast: ${{ contains(github.event.pull_request.labels.*.name, 'agentx-fast') }}
 
     sweep-multi-node-agentic:
         needs: [setup, canary-select, canary-sweep]
@@ -685,6 +690,7 @@ jobs:
             duration: ${{ matrix.config.duration }}
             run-eval: false
             scenario-type: agentic-coding
+            agentx-fast: ${{ contains(github.event.pull_request.labels.*.name, 'agentx-fast') }}
 
     sweep-evals:
         needs: [setup, canary-select, canary-sweep]
@@ -1097,7 +1103,8 @@ jobs:
               github.event.label.name == 'full-sweep-fail-fast' ||
               github.event.label.name == 'full-sweep-fail-fast-no-canary' ||
               github.event.label.name == 'all-evals' ||
-              github.event.label.name == 'evals-only'
+              github.event.label.name == 'evals-only' ||
+              github.event.label.name == 'agentx-fast'
             )
         runs-on: ubuntu-latest
         permissions:

.github/workflows/stage-results-callback.yml

@@ -66,9 +66,17 @@ jobs:
               ? [
                   `@${requestedBy} staged run ${runId}: ${chartUrl}`,
                   '',
-                  `This shared staging slot remains available until the next \`/stage-results\` request. [Staging workflow](${appRunUrl})`,
+                  `This run remains available across future \`/stage-results\` requests. Staging the same run ID again updates its staged data. [Staging workflow](${appRunUrl})`,
+                  '',
+                  `@${requestedBy} 已将运行 ${runId} 发布到预发布环境：${chartUrl}`,
+                  '',
+                  `后续的 \`/stage-results\` 请求不会移除此运行；再次发布相同的运行 ID 会更新其预发布数据。[预发布工作流](${appRunUrl})`,
                 ].join('\n')
-              : `@${requestedBy} staging run ${runId} failed. [Inspect the staging workflow](${appRunUrl}).`;
+              : [
+                  `@${requestedBy} staging run ${runId} failed. [Inspect the staging workflow](${appRunUrl}).`,
+                  '',
+                  `@${requestedBy} 预发布运行 ${runId} 失败。[查看预发布工作流](${appRunUrl})。`,
+                ].join('\n');
 
             if (commentId) {
               await github.rest.issues.updateComment({

.github/workflows/stage-results.yml

@@ -153,14 +153,20 @@ jobs:
                 name === 'eval_results_all' ||
                 name.startsWith('bmk_agentic_'),
             );
-            const validateRun = async (candidate) => {
+            const validateRun = async (candidate, allowHistoricalAssociation = false) => {
               if (
                 candidate.path !== '.github/workflows/run-sweep.yml' ||
                 candidate.event !== 'pull_request'
               ) {
                 return { valid: false, reason: 'not a pull-request run of run-sweep.yml' };
               }
-              if (!pullCommitShas.has(candidate.head_sha)) {
+              const isAssociatedWithPull = candidate.pull_requests?.some(
+                (candidatePull) => candidatePull.number === context.issue.number,
+              );
+              if (
+                !pullCommitShas.has(candidate.head_sha) &&
+                !(allowHistoricalAssociation && isAssociatedWithPull)
+              ) {
                 return { valid: false, reason: 'its head commit is not in the PR commit list' };
               }
               if (fullSweepLabelsAt(candidate.created_at).length === 0) {
@@ -169,13 +175,19 @@ jobs:
                   reason: 'it was not created while a full-sweep label was applied',
                 };
               }
-              if (candidate.status !== 'completed' || candidate.conclusion !== 'success') {
+              const stageableConclusions = new Set(['success', 'cancelled', 'failure']);
+              if (
+                candidate.status !== 'completed' ||
+                !stageableConclusions.has(candidate.conclusion)
+              ) {
                 return {
                   valid: false,
                   reason: `it is ${candidate.status}/${candidate.conclusion}`,
                 };
               }
 
+              // Cancelled or failed sweeps may still contain useful partial results. The
+              // artifact checks below keep empty or metadata-only runs ineligible.
               const artifactNames = await usableArtifactNames(candidate.id);
               if (!artifactNames.includes('changelog-metadata')) {
                 return { valid: false, reason: 'it has no unexpired changelog-metadata artifact' };
@@ -195,7 +207,7 @@ jobs:
                 run_id: Number(requestedRunId),
               });
               run = response.data;
-              const validation = await validateRun(run);
+              const validation = await validateRun(run, true);
               if (!validation.valid) {
                 core.setFailed(`Run ${requestedRunId} cannot be staged because ${validation.reason}`);
                 return;
@@ -222,7 +234,7 @@ jobs:
               }
               if (!run) {
                 core.setFailed(
-                  `No successful completed run-sweep.yml run with unexpired staging artifacts was found for PR #${context.issue.number}`,
+                  `No stageable completed run-sweep.yml run with unexpired staging artifacts was found for PR #${context.issue.number}`,
                 );
                 return;
               }
@@ -244,7 +256,11 @@ jobs:
         with:
           github-token: ${{ github.token }}
           script: |
-            const body = `@${process.env.REQUESTED_BY} staging [run ${process.env.RUN_ID}](${process.env.RUN_URL}). The shared staging slot is serialized; a completion link will be posted here.`;
+            const body = [
+              `@${process.env.REQUESTED_BY} staging [run ${process.env.RUN_ID}](${process.env.RUN_URL}). Existing staged runs will be preserved; a completion link will be posted here.`,
+              '',
+              `@${process.env.REQUESTED_BY} 正在将[运行 ${process.env.RUN_ID}](${process.env.RUN_URL})发布到预发布环境。已有的预发布运行会继续保留；完成后将在此处提供链接。`,
+            ].join('\n');
             const comment = await github.rest.issues.createComment({
               owner: context.repo.owner,
               repo: context.repo.repo,

AGENTS.md

@@ -95,7 +95,7 @@ PRs do not run the sweep automatically - `run-sweep.yml` is gated on a primary s
 - `full-sweep-fail-fast` - runs the full intermediate concurrency sweep behind the same sequential single-node canary gate as `full-sweep-enabled` (so a globally broken change burns one job, not the whole fan-out), and with `strategy.fail-fast` enabled on every matrix: the first failure in a matrix cancels that matrix's remaining jobs. Fail-fast is matrix-scoped, so the other matrices (1k1k vs 8k1k vs agentic vs evals) keep running and self-terminate on their own first failure; their completed results remain valid. The failing job keeps its red *failure* conclusion and the run concludes failed. **This is the strongly recommended default for full sweeps** (image bumps, recipe changes, bring-up) - a failure means the rest of that matrix is wasted GPU time. Caveat: one flaky job kills its matrix's in-flight results; if that repeatedly bites a specific config, fall back to `full-sweep-enabled` for that PR.
 - `full-sweep-fail-fast-no-canary` - same as `full-sweep-fail-fast` but without the canary gate: all matrices fan out immediately. Use when the canary is flaky or not representative of the affected configuration but you still want per-matrix fail-fast.
 
-`all-evals` and `evals-only` are optional modifier labels. Combine either or both with one primary sweep label. `all-evals` expands eval selection to every generated fixed-sequence configuration without changing throughput. `evals-only` suppresses throughput while keeping the default eval subset; combining both runs every eval and no throughput. `all-evals` remains eligible for artifact reuse when paired with an eligible full-sweep label. Runs with `evals-only`, including runs with both modifiers, are not eligible.
+`all-evals`, `evals-only`, and `agentx-fast` are optional modifier labels. Combine them with one primary sweep label. `all-evals` expands eval selection to every generated fixed-sequence configuration without changing throughput. `evals-only` suppresses throughput while keeping the default eval subset; combining both eval modifiers runs every eval and no throughput. `agentx-fast` applies the 5-minute warmup and 20-minute profiling preset to single- and multi-node AgentX throughput jobs only; fixed-sequence and eval jobs are unchanged. `all-evals` remains eligible for artifact reuse when paired with an eligible full-sweep label. Runs with `evals-only` or `agentx-fast` are not eligible for artifact reuse.
 
 **The sweep does not trigger while the PR has merge conflicts.** Even with a sweep label applied, the `run-sweep.yml` workflow will not start until the PR cleanly merges into main — a stale claude/* or update-* branch with a `perf-changelog.yaml` conflict (the common case) will sit in NO_SWEEP / NO_SUCCESS until rebased. Resolution recipe is documented in `KLAUD_DEBUG.md §1.1`: `git merge origin/main`, then `git checkout origin/main -- perf-changelog.yaml`, then re-append the PR's own changelog entry at the tail. Don't 3-way merge `perf-changelog.yaml`; whitespace edits silently re-trigger the deletion check.