Commit Graph

9 Commits

Author SHA1 Message Date
Jesse Gross
22d6c817f8 mlxrunner: fuse top-P and top-K into a single sort pass
When both filters are active, avoid paying for a full sort in top-P
and a partial sort in top-K. Single-filter paths are unchanged.
Improves generation throughput on gemma4:e4b by 1.5%.
2026-04-20 17:43:00 -07:00
Jesse Gross
ca01373b28 mlxrunner: use MaxAxis in the min-P sampler
One reduction op instead of Argmax + TakeAlongAxis.
2026-04-20 17:43:00 -07:00
Jesse Gross
24e038d56a mlxrunner: add logprobs support
Match the ollamarunner and OpenAI semantics: raw, full-vocab log-softmax
with the top-K ranked by probability. Skipped on the GPU when the request
doesn't ask for logprobs so decode doesn't pay for it otherwise.
2026-04-20 17:43:00 -07:00
Daniel Hiltgen
ff23dd343f mlx: apply repeat penalties in sampler (#15631) 2026-04-18 07:49:38 -07:00
Jesse Gross
6f8ddbb26b mlxrunner: fix Slice(0, 0) returning full dimension instead of empty
Slice used cmp.Or to resolve a zero stop value to the dimension size,
intended to support open-ended slices like a[i:]. This made Slice(0, 0)
indistinguishable from Slice(), so any slice with a zero stop would
silently include the entire dimension instead of being empty.

Replace cmp.Or with an explicit End sentinel and resolve negative
indices against the dimension size, matching Python/PyTorch semantics.
2026-03-18 16:06:33 -07:00
Daniel Hiltgen
10e51c5177 MLX: add header vendoring and remove go build tag (#14642)
* prefer rocm v6 on windows

Avoid building with v7 - more changes are needed

* MLX: add header vendoring and remove go build tag

This switches to using a vendoring approach for the mlx-c headers so that Go
can build without requiring a cmake first.  This enables building the new MLX
based code by default.  Every time cmake runs, the headers are refreshed, so we
can easily keep them in sync when we bump mlx versions.  Basic Windows
and Linux support are verified.

* ci: harden for flaky choco repo servers

CI sometimes fails due to choco not actually installing cache.  Since it just speeds up the build, we can proceed without.

* review comments
2026-03-09 17:24:45 -07:00
Patrick Devine
d126467d5d x/mlxrunner: replace sampler interface chain with single stateful Sampler (#14652)
- Collapse MLX sampling state into a single sample.Sampler struct (options + history).
- Replace interface-based sampler chain (TopP, TopK, penalty, etc.) with function-based transforms.
- Update request/pipeline wiring to use *sample.Sampler, seed history from prompt tokens, and append generated tokens each step.
- Implement top_p, min_p, repeat_penalty, and frequency_penalty
2026-03-07 17:50:57 -08:00
Patrick Devine
d18dcd7775 mlxrunner fixes (#14247)
* load glm4_moe_lite from the mlxrunner

* fix loading diffusion models

* remove log lines

* fix --imagegen flag
2026-02-13 22:30:42 -08:00
Patrick Devine
44bdd9a2ef Add MLX runner with GLM4-MoE-Lite model support (#14185)
This change adds a new MLX based runner which includes:

  * Method-based MLX bindings
  * Subprocess-based MLX runner (x/mlxrunner)
  * KV cache with tree management
  * A basic sampler

The GLM4-MoE-Lite model has been ported to use the new bindings.

---------

Co-authored-by: Michael Yang <git@mxy.ng>
2026-02-10 14:57:57 -08:00