fix(apr-cli serve): try Q4K pool-allocator path before generic GPU dispatch (closes #471)#1663
Merged
Merged
Conversation
…spatch (closes #471) `apr serve run <model.apr> --gpu` was routing all APR models through the generic `OwnedQuantizedModel::from_apr()` path, which does per-tensor `cuMemAlloc`. For large Q4K MoE models (e.g. qwen3-coder-30b — 17 GB, 18,867 tensors) that path hangs indefinitely during weight upload. The ALB-095/098 `start_apr_q4k_server_gpu` function (already implemented via `spawn_apr_q4k_inference_thread` → `upload_apr_q4k_weights`'s pool allocator → single `cuMemAlloc` for all tensors) was compiled in but NEVER dispatched from `start_apr_server`. `realizar serve --gpu` uses this path and loads the same 30B MoE in ~12s. Fix: in `start_apr_server`, call `start_apr_q4k_server_gpu` first when `--gpu` is requested. If the model isn't Q4K, `parse_apr_q4k_config` or `upload_apr_q4k_weights` returns an error and we fall through to the existing generic GPU path (which still handles dequantized + small non-Q4K APR fine). Without the `cuda-batch` feature the Q4K function is a stub that errors immediately, so non-batch builds keep their existing fallback chain. Verified: - `cargo check -p apr-cli --features cuda` builds clean - `cargo check -p apr-cli --features cuda-batch` builds clean
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Closes #471. `apr serve run <model.apr> --gpu` hung indefinitely on 17 GB / 18k-tensor Q4K MoE models (qwen3-coder-30b). `realizar serve --gpu` loaded the same file in ~12 s using `spawn_apr_q4k_inference_thread`'s pool allocator.
Root cause: `start_apr_q4k_server_gpu` (already implemented in `handlers_include_01.rs:96`) was compiled in but NEVER called from `start_apr_server`.
Fix
Dispatch `start_apr_q4k_server_gpu` before `start_apr_server_gpu` when `--gpu` is requested:
```rust
match start_apr_q4k_server_gpu(model_path, config) {
Ok(()) => return Ok(()),
Err(e) => { eprintln!("[GH-471] declined: {e}"); }
}
match start_apr_server_gpu(model_path, config) { /* existing fallback */ }
```
The Q4K path's `parse_apr_q4k_config` + `upload_apr_q4k_weights` already validate the model shape, so passing non-Q4K APRs through it fails cleanly. The `cuda`-only stub (no `cuda-batch`) errors immediately, preserving existing behavior for non-batch builds.
Test plan
🤖 Generated with Claude Code