p4: record Phase 52 blocked state (driver opt-out, Needle not loadable, MockModel harness run)

swadhinbiswas committed 3 months ago

aacbf03

1 parent 6b14fa4

commit aacbf03

1 file changed

+13 -0

Files

Showing 1 changed file+13-0

▼PLAN.md

+13-0

@@ -998,6 +998,19 @@ k8s guides, OpenAPI.
998	998	supports CUDA >= 13.3, OR install CUDA toolkit 13.0. Setting
999	999	`CUDARC_CUDA_VERSION=13000` alone is insufficient because nvcc still emits
1000	1000	13.3 PTX.
	1001	+- 2026-07-18 update: A `nvidia-utils 610.43.03` + `linux-cachyos-nvidia-open 7.1.3-2`
	1002	+ driver (CUDA 13.3 runtime) is available in the repos and matches the running
	1003	+ kernel `7.1.3-2-cachyos`. User chose not to perform the driver upgrade
	1004	+ (system change + reboot). Separately, the Needle 26M encoder-decoder weights
	1005	+ were downloaded to `~/needle-bucket` (safetensors/pkl) but are NOT loadable by
	1006	+ `justapi-inference` (which only supports Llama-family GGUF/safetensors), so
	1007	+ they do not unblock this gate. As a non-gate validation, `justapi-gpu-bench`
	1008	+ was run with `MockModel` (CPU, no `--model-path`) — both naive and
	1009	+ SchedulerEngine paths execute; the 813k tok/s "naive" figure is mock-only and
	1010	+ does NOT satisfy the real-GPU acceptance criteria. Phase 52 remains 🔴 blocked
	1011	+ on the CUDA driver fix + a Llama-family GGUF (real file is
	1012	+ `~/models/tinyllama/llama-3.2-1b-q4_k_m.gguf`; the `tinyllama-1.1b.q4_k_m.gguf`
	1013	+ is a 29-byte placeholder).
1001	1014
1002	1015	### Phase 53: HTTP/3 (QUIC) Transport (Status: ⚠️ reverted — dead code removed 2026-07-15)
1003	1016	- Objective: Serve the same Python routes over HTTP/3 (QUIC) in addition to
1004	1017