support dspark - #121
Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 3a79930297
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| CudaDriverApi.cuMemHostAlloc(out rt.PinnedCap, | ||
| new UIntPtr((ulong)((long)_m.NUbatch * feat * 4L)), 0x1 /*PORTABLE*/).ThrowOnError(); |
There was a problem hiding this comment.
When a direct-CUDA model is disposed without terminating the process, this cuMemHostAlloc allocation is never passed to cuMemFreeHost; Dispose() only releases the boundary and token buffers, and the newly allocated _pinnedLogits buffer is omitted as well. PinnedCap can occupy roughly NUbatch * target-layer-count * hidden-size * 4 bytes (tens of MiB), so model reload cycles leak page-locked RAM until subsequent CUDA host allocations or model loads fail. Release both new pinned buffers in Dispose().
Useful? React with 👍 / 👎.
| var turnDecoder = new MtpSpeculativeDecoder(turnSpecModel, turnSpecModel.MtpDraftBlockSize) | ||
| { | ||
| MinDraftProb = 0.35f, |
There was a problem hiding this comment.
When --multi-turn-jsonl is combined with --spec-draft-n-max or --spec-draft-conf-min, both parsed values are ignored: RunMultiTurnTest receives neither option and this decoder always uses the full trained block size and a hard-coded 0.35f threshold. This makes multi-turn performance experiments silently run with different settings than requested; pass the CLI values into this path and apply them as the one-shot path does.
Useful? React with 👍 / 👎.
Engine comparison — TensorSharp vs llama.cpp (PR smoke)No report artifact was produced — the benchmark failed before generating results (see the workflow logs). |
1 participant
No description provided.