feat: upgrade the op collector to vLLM 0.30.0, SGLang 0.5.21 and TRT-LLM 1.3.0rc29

.github/application-test-inventory.json

@@ -9,6 +9,7 @@
   "optional_collection_skips": {
     "tests/unit/collector/test_dsv4_megamoe_workload.py": "real torch required for tensor operations",
     "tests/unit/collector/test_helper_moe_distribution.py": "real torch required for tensor operations",
-    "tests/unit/collector/test_sglang_moe_ep_routing.py": "real torch required for tensor operations"
+    "tests/unit/collector/test_sglang_moe_ep_routing.py": "real torch required for tensor operations",
+    "tests/unit/collector/test_vllm_attn_kv_cache_layout.py": "real torch required for tensor operations"
   }
 }

benchmarks/evidence/collector/op-probe-2026-10/README.md

@@ -0,0 +1,37 @@
+<!--
+SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+SPDX-License-Identifier: Apache-2.0
+-->
+
+# Op-probe findings (2026-09 to 2026-10)
+
+This non-packaged bundle holds the serving observations that the collector
+and generator changes for the vLLM 0.30.0, SGLang 0.5.21 and TensorRT-LLM
+1.3.0rc29 pins cite as provenance. The op-probe harness (`opharness`) produced
+them by booting the real serving framework per model and recording the
+kernels and backends it selected. The harness lands separately.
+
+Citations in the code name either a finding key or a results file below. Grep
+a finding key, such as `sm100_probe_failures_recheck_2026_10_04`, to reach its
+statement.
+
+Every file here is an excerpt: `findings.yaml` keeps the 20 findings the code
+cites, and each results file keeps `_meta` plus the models the citing comments
+name. `_meta.summary` still counts the full matrix. The complete archive lands
+with the op-probe harness.
+
+## Files
+
+- `findings.yaml`: curated conclusions, one entry per key. Each statement gives
+  its own versions, dates and hardware inline, and `applies_to` lists the
+  models it was observed on. Evidence paths inside statements refer to
+  workspace bundles that are not committed.
+- `results/sm100/sglang-0.5.21.yaml`: B200 boot-and-run verdict per model on
+  SGLang 0.5.21 under the default generator render.
+- `results/retests/sm100/sglang-0.5.21.moe_auto.yaml`: the B200 SGLang 0.5.21
+  retest of the MoE models under an explicit `--moe-runner-backend auto`.
+- `results/sm100/trtllm-1.3.0rc29.yaml`: B200 boot-and-run verdict per model on
+  TensorRT-LLM 1.3.0rc29.
+
+Findings describe framework releases and rot with them. Re-verify them on the
+next version bump instead of extending them.

benchmarks/evidence/collector/op-probe-2026-10/results/retests/sm100/sglang-0.5.21.moe_auto.yaml

@@ -0,0 +1,78 @@
+# Excerpt (2026-10-30): _meta plus the models collector comments cite; _meta.summary
+# counts the full matrix, which lands with the op-probe harness.
+_meta:
+  platform: b200_sm100
+  framework: sglang
+  version: 0.5.21
+  what: the MoE cells of results/sm100/<fw>-<ver>.yaml re-probed with the generator's hardware moe_backend fact removed from the rendered config; the as-rendered cells fail with cause 'generator fact rejected
+    by framework'. This file is the identity the framework picks on sm100 when it is allowed to choose.
+  override: --moe-runner-backend deepep_moe (hardware.yaml b200) dropped -> 0.5.21 AUTO
+  summary:
+    pass+override: 72
+    fail: 3
+  evidence: archive/retests/moe_auto/{raw,*.engine.yaml} in the B200 workspace (umbriel-b200-040, 2026-10-04)
+results:
+  deepseek-ai/DeepSeek-V4-Flash:
+    verdict: pass+override
+    attention: DeepseekV4AttnBackend
+    moe_runner: FLASHINFER_MXFP4
+    moe_kernel_families:
+    - trtllm_gen_moe
+    moe_kernels_seen: 4
+    topology: tp1
+  deepseek-ai/DeepSeek-V4-Flash-0731:
+    verdict: pass+override
+    attention: DeepseekV4AttnBackend
+    moe_runner: FLASHINFER_MXFP4
+    moe_kernel_families:
+    - trtllm_gen_moe
+    moe_kernels_seen: 4
+    topology: tp1
+  deepseek-ai/DeepSeek-V4-Pro:
+    verdict: pass+override
+    attention: DeepseekV4AttnBackend
+    moe_runner: FLASHINFER_MXFP4
+    moe_kernel_families:
+    - trtllm_gen_moe
+    moe_kernels_seen: 4
+    topology: tp1
+  deepseek-ai/DeepSeek-V4-Pro-0813:
+    verdict: pass+override
+    attention: DeepseekV4AttnBackend
+    moe_runner: FLASHINFER_MXFP4
+    moe_kernel_families:
+    - trtllm_gen_moe
+    moe_kernels_seen: 4
+    topology: tp1
+  deepseek-ai/DeepSeek-V4.1-Flash:
+    verdict: pass+override
+    attention: DeepseekV4AttnBackend
+    moe_runner: FLASHINFER_MXFP4
+    moe_kernel_families:
+    - trtllm_gen_moe
+    moe_kernels_seen: 4
+    topology: tp1
+  moonshotai/Kimi-K2.5:
+    verdict: pass+override
+    attention: TRTLLMMLABackend
+    moe_runner: FLASHINFER_TRTLLM
+    moe_kernel_families:
+    - trtllm_gen_moe
+    moe_kernels_seen: 5
+    topology: tp1
+  openai/gpt-oss-120b:
+    verdict: pass+override
+    attention: TRTLLMHAAttnBackend
+    moe_runner: FLASHINFER_MXFP4
+    moe_kernel_families:
+    - trtllm_gen_moe
+    moe_kernels_seen: 5
+    topology: tp1
+  openai/gpt-oss-20b:
+    verdict: pass+override
+    attention: TRTLLMHAAttnBackend
+    moe_runner: FLASHINFER_MXFP4
+    moe_kernel_families:
+    - trtllm_gen_moe
+    moe_kernels_seen: 5
+    topology: tp1

benchmarks/evidence/collector/op-probe-2026-10/results/sm100/sglang-0.5.21.yaml

@@ -0,0 +1,34 @@
+# Excerpt (2026-10-30): _meta plus the models collector comments cite; _meta.summary
+# counts the full matrix, which lands with the op-probe harness.
+_meta:
+  platform: b200_sm100
+  framework: sglang
+  version: 0.5.21
+  verdicts: pass = plain `cli generate` output boots+runs; pass+custom = boots with facts-derived extra generate args; fail = root-caused (see results/findings.yaml)
+  summary:
+    pass: 102
+    pass+custom: 6
+    fail: 16
+  golden_facts_not_applied: 0
+results:
+  moonshotai/Kimi-K2.5:
+    verdict: pass
+    evidence: current
+    attention: TRTLLMMLABackend
+    moe: CompressedTensorsFusedMoE->trtllm_gen_moe
+    kv_allocated: auto
+    topology: tp1
+  openai/gpt-oss-120b:
+    verdict: pass
+    evidence: current
+    attention: TRTLLMHAAttnBackend
+    moe: Mxfp4MoE->trtllm_gen_moe
+    kv_allocated: auto
+    topology: tp1
+  openai/gpt-oss-20b:
+    verdict: pass
+    evidence: current
+    attention: TRTLLMHAAttnBackend
+    moe: Mxfp4MoE->trtllm_gen_moe
+    kv_allocated: auto
+    topology: tp1

benchmarks/evidence/collector/op-probe-2026-10/results/sm100/trtllm-1.3.0rc29.yaml

@@ -0,0 +1,17 @@
+# Excerpt (2026-10-30): _meta plus the models collector comments cite; _meta.summary
+# counts the full matrix, which lands with the op-probe harness.
+_meta:
+  platform: b200_sm100
+  framework: trtllm
+  version: 1.3.0rc29
+  verdicts: pass = plain `cli generate` output boots+runs; pass+custom = boots with facts-derived extra generate args; fail = root-caused (see results/findings.yaml)
+  summary:
+    pass: 78
+    pass+custom: 0
+    fail: 46
+  golden_facts_not_applied: 0
+results:
+  moonshotai/Kimi-K2.5:
+    verdict: fail
+    cause: quant parser gap
+    error: 'TypeError: ''NoneType'' object is not subscriptable'

benchmarks/evidence/collector/pr244/README.md

@@ -6,8 +6,17 @@ SPDX-License-Identifier: Apache-2.0
 # Collection audit evidence
 
 This non-packaged bundle verifies the retained vLLM 0.25.0 collection outcomes
-and the 85 shipped Parquet tables. It contains case IDs, status vectors,
-checksums, and table metadata, not diagnostic logs or machine paths.
+and the 80 Parquet tables still shipped from it (85 originally). It contains case
+IDs, status vectors, checksums, and table metadata, not diagnostic logs or
+machine paths.
+
+The five `dsv4_hca_attn_module_perf` tables (one per system) were removed with
+the hca_attn sidecar's SDK reader in #649. The case-outcome ledger keeps their
+reported hashes, which record what was collected. The audit's `RETIRED_TABLES`
+excludes them from the inventory check. The manifest was regenerated: it
+drops their entries and carries the new hashes of the five 0.25.0
+`collection_meta.yaml` sidecars that listed them. Every other table entry is
+unchanged.
 
 Run from the repository root in the development environment: