test: validate four ATOM models over HTTP with device and host cache

opt_prime/examples/pd/http_matrix/.research/events.jsonl

@@ -1 +1,2 @@
 {"id": "b61f99ae005a4258aa7f44d8edeca28d", "ts": "2026-10-23T18:08:44+09:00", "task": "Four-model HTTP matrix (aicomp #211)", "result": "All4 GPU models pass HTTP concurrency4/8, cancel/disconnect/drain/reuse checks. No-cacheBF16 and warm-cacheFP32 each48/48 concurrent outputs exact. Warm-cacheBF16 has31/48 raw equality; actualQwenHTTP tie explains observed branch. Warm-state cancellation isolates KONI cold-cache difference.", "next": "Publish GPUHTTP evidence; continue ATOM HTTP and RNGD routes. Preserve numerical/quality and true-overlap limitations.", "state": "active", "done": false, "blocker": "", "metrics": {}, "params": {"models": "4", "slots": "4", "concurrency": "4,8", "conditions": "bf16_no_cache,bf16_cache,fp32_cache"}, "artifacts": ["evidence/summary.json"], "experiments": ["experiments/2026-10-23-http-matrix211.md"], "labels": ["benchmark"], "review": true, "todos": [], "note": "", "agent": "codex", "host": "GPU host", "branch": "main", "commit": "351d39d-dirty"}
+{"id": "3563031e43c8478e951fff82f81bfe94", "ts": "2026-10-23T18:38:18+09:00", "task": "Four-model serving matrix (aicomp #201)", "result": "ATOM HTTP: all four models pass 128/2048/8192-token lifecycle checks; cache-off concurrent48/48 exact. Both Qwen device/host cache cases pass12/12 each. SIGPIPE candidate suppresses native disconnect backtraces6 to0 with unchanged outputs;14 unit/integration tests pass.", "next": "Publish ATOM HTTP matrix and SIGPIPE fix; continue missing RNGD and hybrid-state routes.", "state": "active", "done": false, "blocker": "", "metrics": {}, "params": {"backend": "atom", "input_tokens": "128,2048,8192", "output_tokens": "32", "hardware_batch": "1"}, "artifacts": ["evidence/atom-http-qwen3-4b-host-fixed.json"], "experiments": ["experiments/2026-10-23-atom-http211.md"], "labels": ["experiment", "bug"], "review": false, "todos": [], "note": "", "agent": "codex", "host": "RNGD+ATOM host", "branch": "exp/211-atom-http", "commit": "working candidate"}

opt_prime/examples/pd/http_matrix/README.md

@@ -1,4 +1,4 @@
-# Four-model GPU HTTP validation (#211/#212)
+# Four-model GPU and ATOM HTTP validation (#211/#212)
 
 Run `validate.py --model MODEL --output result.json` under the engine environment with a free CUDA device. `--prefix-cache` enables the separate cache case; `--dtype float32` provides the numerical control. Offline immutable model snapshots are required. The default configuration is four paged execution slots, context9216, page64, prefill chunk256, inputs128/2048/8192 and32 fixed output tokens. This exercises the real HTTP service factory and sockets, not a mocked handler.
 
@@ -9,3 +9,11 @@ Run `validate.py --model MODEL --output result.json` under the engine environmen
 See experiments/2026-10-23-http-matrix211.md and evidence/summary.json. Evidence includes the initial failed KONI cold-baseline cancellation assertion. Earlier runs preceded the warm-state cancellation control and some metadata additions; their commands/configurations are recorded in the experiment. Public copies normalize local model-directory labels in metrics only; raw measurements remain in the research workspace.
 
 This GPU experiment does not complete ATOM/RNGD HTTP or PD coverage, long-context RNGD support, hybrid NPU cache restoration, broader semantic quality or true compute/transfer overlap.
+
+## ATOM extension
+
+Use `--backend atom` in an isolated allocation exposing exactly one ATOM. Existing artifacts must use B1/context9216 and match `--atom-kv-blocks` (default3; KONI8B requires1). The two Qwen models support `--prefix-cache`; add `--host-kv-gib 2` to exercise the host tier and require positive observed host hits. Artifact precision is fixed; float32 is rejected for ATOM. Set OPT_PRIME_CACHE to the task-owned artifact cache. SDK/device bootstrap is deployment-specific and must map the allocated physical device to runtime-visible0 when required.
+
+Inputs128/2048/8192 and clients4/8 are queued over one ATOM execution lane, not hardware batching. All four cache-off runs pass lifecycle checks and48/48 concurrent/serial comparisons. Both Qwen device/host cache cases pass12/12 each. Gemma requires the causal-prefill SDK graph correction and #210 hybrid local-table/mask fix; its prefix/host/PD restoration remains unsupported. KONI's one-block configuration has no spare prefix block.
+
+See experiments/2026-10-23-atom-http211.md and evidence/atom-http-*.json. Files ending `fixed` include #214's HTTP MSG_NOSIGNAL fix; the Qwen0.6 matched rerun preserves all output sequences and eliminates the observed native disconnect backtraces. This does not establish all RNGD/heterogeneous routes or real transfer overlap.

opt_prime/examples/pd/http_matrix/STATUS.md

@@ -1,3 +1,3 @@
 # HTTP model matrix
 
-State: review; goal remains active. Four GPU models pass no-cache lifecycle checks and48/48 concurrent equality; BF16 warm cache has raw numerical differences. FP32 warm cache passes lifecycle and48/48 concurrent equality. Actual Qwen HTTP trace identifies a BF16 logit tie. Next: ATOM HTTP configurations and missing RNGD routes; broader semantic inputs and cache/host-tier ownership checks.
+State: review; full goal remains active. GPU evidence remains as recorded. ATOM four-model cache-off lifecycle passes with48/48 concurrent equality; both Qwen device/host cache cases pass12/12 each. SIGPIPE fix tracked separately in #214. Remaining: RNGD routes, hybrid-state restoration, real transfer overlap and broader semantic testing.