HIP: release cancelled prompt chunks at layer boundaries

docs/receipts/gfx1151-cancel-76aafcf2/README.md

@@ -0,0 +1,74 @@
+# gfx1151 cancellation follow-up
+
+Measured on 2026-10-18, following the unresolved 27B case in [#994](https://github.com/ashhart/TensorFold/pull/994).
+This is a fresh A/B run on public source `76aafcf274e60a5e31a59ce145b2f71f7b366e9d` and the runtime patch
+in `5429554e708fdf7a9e0017301a00eb87a4182acd`. It is not a full release qualification.
+Subsequent source changes shorten comments/docstrings for the lean rule; runtime and gate execution logic are unchanged.
+
+The single GPU is a BOSGAME M5's Radeon 8060S (`gfx1151`). The checkpoint is
+`TensorFold/Qwen3.8-27B-MLX-4bit` at `70ae7fac63274ff2eac54152031433374cb80f2f`.
+All 17 checkpoint files match their previously pinned public Git-blob/LFS digests; see `checkpoint.json`.
+
+## Cancellation result
+
+The same gate runs three trials per build. Each trial separately measures the time from closing a long request's
+socket until the engine's `requests_running` metric reaches zero, then the next short request sent 50 ms behind
+another cancellation. Every long request uses a fresh nonce first, preventing long-prefix reuse.
+The client closes 0.4 seconds after sending it. All twelve cancelled prompts fit within the 16,384-token window
+and end with `reason=client-left`, zero emitted tokens. Actual prompt lengths are in `audit.json`.
+
+| Measurement | Unmodified base | Fixed |
+| --- | ---: | ---: |
+| Idle short baseline, best of three | 1.899 s | 1.913 s |
+| Disconnect to idle metric, three trials | 4.720–4.744 s | 0.032–0.042 s |
+| Next short request, three trials | 6.532–6.548 s | 1.902–1.907 s |
+| Cancelled request lifetime, arrival to end | 5.08–5.14 s | 0.43–0.44 s |
+| Gate: release <=0.5 s and short <=idle+0.5 s | **3/3 failed** | **3/3 passed** |
+
+All fourteen short replies across the two builds have the same complete native token SHA-256:
+`ad11ca37c7fcabaf13c8fc1142956890438c4ca49c58d6f579c3eef5485b77d5`.
+`*-http.jsonl` retains each short request/response and each cancelled request's exact body.
+`*-server.log` independently retains six client-left completions per build.
+Arrival-to-end lifetime is not used as disconnect-to-release time.
+
+The harness's optional `--baseline-27b` also prints the historical v1.0.12 receipt for comparison.
+Those historical numbers are labelled separately; the table above is the fresh unmodified-versus-fixed comparison.
+
+## Correctness and timing controls
+
+An identical uncancelled 14,975-token prompt produces the same full native token SHA-256 on both builds:
+`677caeac8d5e87f26c67e37ff1551667fc3217c4690d7a276204d22c2504f0a1`.
+Prefill takes 79.021 s before and 78.867 s after. This is one timing control, not a broad performance claim.
+The complete requests and replies are in `*-uncancelled.json`.
+
+The unmodified `serve_check.py` then passes its 13 exercised comparisons on the fixed server, including
+concurrent/solo and resumed/fresh checks. It prints 15 successes, but two drafting comparisons are unexercised:
+this checkpoint has no draft head. The raw output retains all fifteen lines without inflating exercised coverage.
+
+The host build passes 292/292 steps; cached results are reused. It reports 126 new passes and two skips, and the
+HIP Qwen family reports 32 passes/one skip. Before the implementation, retaining the old 2,048-row threshold and
+the null cancellation hook makes both new CPU regressions fail. Remaining work in that deliberately failing build
+was stopped once those two failures were captured; it is not a completed full-suite run. See `cpu-*.log`.
+
+## Reproduction and scope
+
+`state.json` retains exact build/server/test commands and both binary SHA-256 values. Use Zig 0.17.0,
+HIP 7.1 and clang 21; the HIP build targets `gfx1151` with explicit compiler/include/device-bitcode paths.
+The server uses `--context 16384 --parallel 4 --prompt-cache-gib 1 --no-thinking --metrics-open` on loopback.
+Run from each source version, using the same pinned checkpoint:
+
+```sh
+python3 zig/tests/qwen35/serve_cancel_release.py http://127.0.0.1:8092 amd-cancel-27b \
+  --runs 3 --server-log SERVER_LOG --record HTTP_JSONL --baseline-27b
+python3 zig/tests/qwen35/serve_check.py http://127.0.0.1:8092 amd-cancel-27b 64
+```
+
+Only single-rank HIP Qwen prompt passes of at least 512 rows now poll cancellation at layer boundaries.
+Shorter passes still finish their current chunk; tensor-parallel groups retain between-chunk cancellation so
+ranks do not diverge at a collective. This run does not qualify every prompt cut, other models, multi-GPU,
+drafting, or all release gates. It does not implement DFlash2 or resolve the separate benchmark-publication work.
+
+This directory is a separate sanitized publication copy. Machine-local paths/hostname are replaced with
+placeholders and trailing whitespace normalized. Request/response values and native hashes are unchanged.
+`publication.json` records source/published hashes; `SHA256SUMS` covers every published file except itself.
+Original raw evidence and service-control state remain private and unchanged.

docs/receipts/gfx1151-cancel-76aafcf2/SHA256SUMS

@@ -0,0 +1,21 @@
+4dfa053b09d0f5584088538097e0094a19bda027f59f0363bc8c747aa0f50e82  README.md
+d4f77f3605d3ede0aeef5757c360bba26823b1baa46c592ba7de3c1fb5f34392  apply-fix.log
+4ba097ef30465f860f9d123ba3cae1281f3cef8f71658522cd9c806c357e8a3c  audit.json
+01ba4719c80b6fe911b091a7c05124b64eeece964e09c058ef8f9805daca546b  base-build.log
+5127ed813d5e1490682f5859fef40327aae50cdc7ac44add58d65afe292d11b4  base-cancel.log
+694402e41dc987834b2314fa71be2ed4386496730f72a60bbb7d2702e21e33cc  base-http.jsonl
+376990664a4a2514fd468950e949c221b209d2c26e15e2636c7b1c3aa7da4ce4  base-server.log
+2a8f77d9ca5990ecbf3b54812af25bcde7a35123e4e67473e97e570054e56561  base-uncancelled.json
+4c3cd0481e429da05b394a479be92ecc3017143d7aaa9224348b55db3896626e  checkpoint.json
+24adf478be7254d6d16f95925dd4215dd2858170193713ad4108e0811dc11221  cpu-fixed.log
+307573b7894c6f340ec371e4b088f20e8bec07a5056aa3cea83d6fcdfc024d19  cpu-negative.log
+2381a774b0b278f6ffc478ebf0d21592d466a1211b0c839221aa5c6cf40390a7  environment.json
+01ba4719c80b6fe911b091a7c05124b64eeece964e09c058ef8f9805daca546b  fixed-build.log
+dc3344269ef666913603f710f9491e3f047bdd4b02c4025ecbd2f46ecd4d446e  fixed-cancel.log
+e64d7c189bc2aee370b1369dfa652e164585439b9acafa9529f512e05726ff2c  fixed-http.jsonl
+075c7d1ef54c7124576d094fc0badccc58565fbdce83bde2c3de9ad2ed961eb8  fixed-serve-check.log
+22aea9113e62618196f70101b230999c897ad11b7b408aa64b506a24fcf43f31  fixed-server.log
+45516e70ae391ccfb654e02b0a94ae741b89c0095df733936d41b33129750bb8  fixed-uncancelled.json
+3ccf8f9783270c302f6709bd7e1418afa9685be26c7c6ed23120f631f6f1d49c  provenance.json
+78281c6f127f352d593298fe89b1e759e16432f4e89eb940f27eec9355289499  publication.json
+984fa52ed6f27a7b79ef034691cd986e2b1ad7986ca2584a9f49a84a793fb0df  state.json

docs/receipts/gfx1151-cancel-76aafcf2/apply-fix.log

@@ -0,0 +1,4 @@
+patching file zig/src/families/qwen3_5/backend/hip/engine/engine.zig
+patching file zig/src/families/qwen3_5/backend/hip/engine/hip_lanes.zig
+patching file zig/src/families/qwen3_5/backend/hip/engine/prefix.zig
+patching file zig/src/families/qwen3_5/hip.zig

docs/receipts/gfx1151-cancel-76aafcf2/audit.json

@@ -0,0 +1,24 @@
+{
+  "cancelled_prompt_tokens": {
+    "base": [
+      14994,
+      14994,
+      14994,
+      14994,
+      14994,
+      14994
+    ],
+    "fixed": [
+      14994,
+      14994,
+      14994,
+      14994,
+      14994,
+      14994
+    ]
+  },
+  "cancelled_prompts_fit_context": true,
+  "short_http_responses": 14,
+  "identical_short_native_hashes": true,
+  "identical_long_native_hashes": true
+}

docs/receipts/gfx1151-cancel-76aafcf2/base-build.log

@@ -0,0 +1 @@
+

docs/receipts/gfx1151-cancel-76aafcf2/base-cancel.log

@@ -0,0 +1,8 @@
+FAIL cancel 0: released 4.73 s after the disconnect (max 0.50); the next short request took 6.55 s (alone 1.90 s, slack 0.50)
+FAIL cancel 1: released 4.74 s after the disconnect (max 0.50); the next short request took 6.53 s (alone 1.90 s, slack 0.50)
+FAIL cancel 2: released 4.72 s after the disconnect (max 0.50); the next short request took 6.54 s (alone 1.90 s, slack 0.50)
+server: 6 cancelled requests ended 5.08 to 5.14 s after they arrived (arrival to end, not disconnect to release)
+1.0.12 receipt 27B: next short 6.49 to 6.51 s, alone 1.90 s, cancelled requests ended 5.04 to 5.07 s after arrival
+this run: next short at most 6.55 s, alone 1.90 s: -0.06 s against the receipt's best
+summary {"model": "amd-cancel-27b", "words": 2200, "drop_after_s": 0.4, "alone_s": 1.899, "slack_s": 0.5, "release_max_s": 0.5, "trials": [{"release_s": 4.73, "after_short_s": 6.548, "running_at_drop": true, "same_text": true, "same_native_tokens": true, "ok": false}, {"release_s": 4.744, "after_short_s": 6.532, "running_at_drop": true, "same_text": true, "same_native_tokens": true, "ok": false}, {"release_s": 4.72, "after_short_s": 6.543, "running_at_drop": true, "same_text": true, "same_native_tokens": true, "ok": false}], "baseline_native_token_sha256": "ad11ca37c7fcabaf13c8fc1142956890438c4ca49c58d6f579c3eef5485b77d5", "lifetime_s": [5.13, 5.1, 5.14, 5.08, 5.12, 5.08], "failed": 3, "receipt_27b": {"after_short_s": [6.49, 6.51], "alone_s": 1.9, "lifetime_s": [5.04, 5.07]}}
+failed 3