← sources

Follow-up: my native Rust + Vulkan Transformer training backend — 14 days later, now 14 parity-verified architectures and full PEFT

source: reddit (reddit:singularity)published 2026-09-30 14:41ingested 2026-10-01 02:41

retained text

Follow-up to my post from about two weeks ago. A lot has changed since then, so I wanted to post an update on where the backend is now.

# Where the green architectures stand

When I posted last time, 7 architectures had verified full training support. Everything is now held to the same strict harness: a pinned local Hugging Face Transformers source tree as the oracle, forward logits + gradients + two full AdamW steps compared, and every named parameter checked again after export.

The hard ceiling is `2e-7` absolute error. No loosening tolerances and no rounding numbers afterward to make the README look better.

**14 architectures pass that gate today**, led by the one I'm probably proudest of:

|Architecture|Scope|
|:-|:-|
||
|**Falcon H1 / H1R**|parallel GQA/RoPE attention + Mamba2 in every layer; full training, full fine-tuning, LoRA, saved modules|
|DeepSeek V4|causal LM|
|Phi-4 Multimodal|text backbone|
|Phi-3|causal LM|
|Kimi K2.5|text backbone|
|Kimi K3 / KimiLinear|hybrid KDA + MLA|
|GPT-OSS|causal LM incl. router bias|
|SmolLM3|mixed RoPE/NoPE + YaRN|
|Qwen2.5 / Qwen3.5 / Qwen4-Exp|dense, DeltaNet, QSA, PLE, MoE|
|Mistral 4, MiniMax M3, Gemma 3/4, MiniMax M2|causal LM|

Worst observed two-step AdamW parameter error across all of them: `1.19e-7`.

Best: `2.6e-8`.

For hardware context, **all of the local Vulkan validation I've been reporting was run on my ASUS ROG Ally Z1 Extreme**, using its AMD RDNA 3 integrated GPU. So the RDNA 3 results here are from that specific machine rather than testing across several different AMD systems.

# The bigger news: PEFT actually works now

In the last post, "LoRA/PEFT-style fine-tuning" was basically one line in a feature list.

It's a real workflow now, and I've verified the full lifecycle:

* **LoRA fine-tuning** with HF-compatible adapter export (`adapter_config.json` / `adapter_model.safetensors`), so adapters can round-trip with the PEFT ecosystem
* `modules_to_save` — full trainable replacements for Linears, RMSNorm/LayerNorm, `lm_head`, and input embeddings, including named-adapter switching and bank isolation. Adapter A leaking into adapter B is explicitly tested for.
* **Exact resume** — adapter weights + AdamW moments + step + dropout RNG state restore bit-identically against an uninterrupted run
* **Merge/unmerge**, disable-adapter base restoration, and multi-adapter loading
* A **parameter-budget flag** that automatically chooses the largest LoRA rank that fits within a requested percentage of the base model
* The CLI **fails closed** if you try to use saved modules on an architecture that hasn't passed its corresponding gate

**32 architecture surfaces across 20 families pass all three PEFT stages** — LoRA, saved modules, and adapter switching — under the same `2e-7` gate, with frozen-base drift exactly `0.0`.

The validation harness also fingerprints the pinned Transformers source alongside my shaders and binaries now, so a qualification run can't silently end up testing against different reference math.

# A small note on the last couple weeks

I didn't get quite as many working days out of the last two weeks as I normally would have. Partway through this I got covid, then when that started to go away, it became a secondary nasty ear infection that ended up perforating my eardrum. I was running fevers around 104°F at one point and eventually went to the hospital, so I lost a few days to that and I'm on antibiotics now.

I'm doing better, though, and still managed to get most of what I wanted finished.

There are still things I want to clean up and expand, but I figured this was a good point to get the current work in front of people rather than holding the update back.

# Same caveats as before

This is deterministic FP32 tiny-model correctness against a reference implementation, which should in theory ensure total mathematical parity for training and finetuning larger models with this backend, however, things are currently bound to FP32 training runs still, I eventually plan to work on MXFP4 weight tying to reduce memory footprints while fine-tuning (plastic parameters will still be trained in FP32 with PEFT in this config)

**"supported text graph" ≠ "the entire multimodal package works natively."**

Unsupported functionality is supposed to fail closed rather than silently falling back to an approximation.

# Repo

[https://github.com/necat101/Hierarchos-Native](https://github.com/necat101/Hierarchos-Native)

* Architecture inventory: `hierarchos-vulkan/README_ARCHITECTURES.md`
* Compatibility/parity record: `hierarchos-vulkan/COMPATIBILITY.md`
* PEFT qualification evidence: `PROGRESS_PEFT_AUDIT.md`
* CLI PEFT guide: `hierarchos-native-cli/README.md`

The hardware I've personally validated this on is **an ASUS ROG Ally Z1 Extreme with its AMD RDNA 3 GPU**.

I'm very interested in criticism, compatibility reports, and especially results from people trying it on other hardware — NVIDIA, Intel, or other AMD GPUs.

I'd also love people to stress-test the PEFT resume/merge paths specifically. That's some of the newest code in the project, so it's probably the most useful area to try to break right now.

document fetched 2026-10-01 00:34 · canonical https://reddit.com/r/singularity/comments/1wu6qta/followup_my_native_rust_vulkan_transformer