Skip to content

Add regression test for trust_remote_code passthrough (#1724) - #1

Open
camparker wants to merge 10 commits into
mainfrom
fix-1724-trust-remote-code
Open

Add regression test for trust_remote_code passthrough (#1724)#1
camparker wants to merge 10 commits into
mainfrom
fix-1724-trust-remote-code

Conversation

@camparker

Copy link
Copy Markdown
Owner

No description provided.

@camparker
camparker force-pushed the fix-1724-trust-remote-code branch from 12da3da to 24e229c Compare July 27, 2026 04:06
lucasnewman and others added 8 commits July 27, 2026 09:28
Adds pfnet/plamo-2.1-2b-vl and plamo-2.1-8b-vl: a Mamba/Attention hybrid
plamo2 language model with a SigLIP so400m vision encoder and an
RMSNorm+Bias MLP projector. The language side is adapted from mlx-lm's
plamo2 with per-layer RoPE base selection (plamo-2.1 uses rope_theta 1e6)
and inputs_embeds support.

Verified against the upstream transformers implementation (bf16, MPS) on
identical inputs: prefill top-1 logits match 6/6 (cos >= 0.9986) for both
2B and 8B, and greedy photo-description outputs match string-for-string.
* Add llm-jp-4-vl (llmjpvl) model support

Adds llm-jp/llm-jp-4-vl-9B-beta: InternVL-style glue (dynamic tiling +
pixel_shuffle + mlp1 projector) over a SigLIP vision encoder and a Llama
language model, built from mlx_vlm's shared components.

Verified greedy-parity with the upstream transformers implementation
(bf16, MPS): 6/6 exact string matches on a Japanese document QA set with
identical pre/post-processing.

* Register llmjpvl in prompt_utils for CLI support

---------

Co-authored-by: Prince Canuma <prince_canuma@Princes-MacBook-Pro.local>
* Fix Qwen2.5-VL chunked MRoPE positions

* Cover chunked MRoPE across Qwen VL models
* Add Kimi K3

* Move chat templating into processor, add tests

---------

Co-authored-by: Prince Canuma <prince.gdt@gmail.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

5 participants