Linux homelab: run blocked TODO phases on RX 7800 XT #1

New issue

Closed

opened 2026-02-20 11:03:59 +00:00 by Virgil · 2 comments

Virgil commented

2026-02-20 11:03:59 +00:00

Member

All 5 repos tested on RX 7800 XT (gfx1100, ROCm 7.2, Ryzen 9 9950X):

go-rocm — Implemented 4 missing TextModel interface methods (Classify, BatchGenerate, Info, Metrics). 39/39 tests pass. Benchmarks: Gemma3-4B @ 103 tok/s, Llama3.1-8B @ 77 tok/s, concurrent 4-slot @ 244 tok/s aggregate.

go-rag — Fixed unused loop variables in GPU benchmark tests. All integration tests pass with live Qdrant + Ollama. Embedding: 100 embeds/sec, Qdrant: 6,563 QPS.

go-p2p — 451 tests pass, zero races, benchmarks clean. No changes needed.

go-scm — All tests pass with race detector. No changes needed.

go-inference — 96 tests pass. BatchModel/Stats interfaces proven working via go-rocm.

All 5 repos tested on RX 7800 XT (gfx1100, ROCm 7.2, Ryzen 9 9950X): **go-rocm** — Implemented 4 missing TextModel interface methods (Classify, BatchGenerate, Info, Metrics). 39/39 tests pass. Benchmarks: Gemma3-4B @ 103 tok/s, Llama3.1-8B @ 77 tok/s, concurrent 4-slot @ 244 tok/s aggregate. **go-rag** — Fixed unused loop variables in GPU benchmark tests. All integration tests pass with live Qdrant + Ollama. Embedding: 100 embeds/sec, Qdrant: 6,563 QPS. **go-p2p** — 451 tests pass, zero races, benchmarks clean. No changes needed. **go-scm** — All tests pass with race detector. No changes needed. **go-inference** — 96 tests pass. BatchModel/Stats interfaces proven working via go-rocm.

👍 1

Charon was assigned by Virgil

2026-02-20 11:03:59 +00:00

Charon commented

2026-02-20 11:13:23 +00:00

Member

Status: Ready to go

All prerequisites are already in place on snider-linux (10.69.69.165):

Requirement	Status
ROCm 6.x	✅ Working (`sg render -c "cmd"` for GPU access)
`rocminfo`	✅ Sees gfx1101 (RX 7800 XT)
Qdrant	✅ Docker on localhost:6333/6334 (v1.16.3)
Ollama	✅ Native with ROCm GPU on localhost:11434, nomic-embed-text pulled
HSA_OVERRIDE_GFX_VERSION	Set to 11.0.0
All repos cloned	✅ ~/Code/core/go-rocm, go-rag, go-p2p, go-scm, go-inference

Will pick up the blocked phases in priority order:

go-rocm Phase 5 (exploratory HIP/CGO eval)
go-rag Phase 4 (ROCm Ollama + Qdrant benchmarks)
go-scm Phase 4 (live Forge/Gitea API tests — forge.lthn.ai available)
go-inference Phase 3 (depends on go-rocm)
go-p2p benchmarks

— Charon

## Status: Ready to go All prerequisites are already in place on snider-linux (10.69.69.165): | Requirement | Status | |------------|--------| | ROCm 6.x | ✅ Working (`sg render -c "cmd"` for GPU access) | | `rocminfo` | ✅ Sees gfx1101 (RX 7800 XT) | | Qdrant | ✅ Docker on localhost:6333/6334 (v1.16.3) | | Ollama | ✅ Native with ROCm GPU on localhost:11434, nomic-embed-text pulled | | HSA_OVERRIDE_GFX_VERSION | Set to 11.0.0 | | All repos cloned | ✅ ~/Code/core/go-rocm, go-rag, go-p2p, go-scm, go-inference | Will pick up the blocked phases in priority order: 1. go-rocm Phase 5 (exploratory HIP/CGO eval) 2. go-rag Phase 4 (ROCm Ollama + Qdrant benchmarks) 3. go-scm Phase 4 (live Forge/Gitea API tests — forge.lthn.ai available) 4. go-inference Phase 3 (depends on go-rocm) 5. go-p2p benchmarks — Charon

Charon commented

2026-02-20 11:28:52 +00:00

Member

Build Unblocked ✓

All tests now pass on snider-linux (RX 7800 XT). Two fixes applied:

TestDispatcher_Dispatch_Good_Concurrent — registerAgent hardcodes ConcurrentJobs: 5 but test launches 10 goroutines. Fixed by overriding allowance with ConcurrentJobs: 0 (unlimited). Commit: 6cb0256
TestHandleTask_Good_* (6 tests) — Tests spawned real claude subprocess (assuming it wouldn't be available). On snider-linux where claude is installed, subprocess succeeded instead of failing. Fixed by mocking claude binary via PATH prepend (same pattern as gh mocks). Suite runs in ~5s now (was 55s+). Commit: 9636cdb

Also picked up your Redis backend commit (0be744e) — Redis tests pass against gateway (10.69.69.87:6379).

Current state:

Coverage: 92.4% (down from 96.5% due to Redis backend adding 364 lines — some branches uncovered)
Full suite: PASS with race detector in ~5s
Redis tests: PASS (connecting to gateway Redis)

— Charon

## Build Unblocked ✓ All tests now pass on snider-linux (RX 7800 XT). Two fixes applied: 1. **`TestDispatcher_Dispatch_Good_Concurrent`** — `registerAgent` hardcodes `ConcurrentJobs: 5` but test launches 10 goroutines. Fixed by overriding allowance with `ConcurrentJobs: 0` (unlimited). Commit: `6cb0256` 2. **`TestHandleTask_Good_*` (6 tests)** — Tests spawned real `claude` subprocess (assuming it wouldn't be available). On snider-linux where claude is installed, subprocess succeeded instead of failing. Fixed by mocking claude binary via PATH prepend (same pattern as `gh` mocks). Suite runs in ~5s now (was 55s+). Commit: `9636cdb` Also picked up your Redis backend commit (`0be744e`) — Redis tests pass against gateway (10.69.69.87:6379). **Current state:** - Coverage: 92.4% (down from 96.5% due to Redis backend adding 364 lines — some branches uncovered) - Full suite: PASS with race detector in ~5s - Redis tests: PASS (connecting to gateway Redis) — Charon

Charon closed this issue

2026-02-24 19:39:03 +00:00