Pin llama.cpp 83e1c41: OPT, CodeGen, GPT-Neo, GPT-J and 48 sibling HF names by bong-water-water-bong · Pull Request #138 · 1bit-MONSTER/engine

bong-water-water-bong · 2026-09-26T14:16:32Z

Moves third_party/llama.cpp from 706ad1a to 83e1c41, the merge of 1bit-MONSTER/llama.cpp#8 and the head of 1bit/hrx-vulkan-patched.

What the pin adds. Four architectures upstream llama.cpp has no model for (for gptj it has only a name), plus sibling spellings of HF classes the runtime already has. I checked each new port against transformers FP32 before merging #8 (teacher-forced top-1, 3 prompts × 32 tokens, one prompt of 340 tokens). Four of them needed fixes first:

model	agreement	fixed on the way
facebook/opt-125m	96/96	tokenizer checksum
EleutherAI/gpt-neo-125m	96/96	no attention scaling, as transformers does
Salesforce/codegen-350M-mono	96/96 (was 0/96)	the (query, value, key) blocks of `qkv_proj`; the `lm_head` bias
tiny-random-GPTJForCausalLM	89/89	`lm_head` bias, mask buffers, tokenizer checksum

I also dropped the BertForCausalLM/BertLMHeadModel aliases: those are causal decoders, and the bert arch is a bidirectional encoder.

Engine changes.

serve: fork_only_arch now covers opt, codegen, gptneo and gptj next to zaya. Checked with the upstream Vulkan server present: OPT-125m and GPT-Neo-125m both answer through 1bit serve --device vulkan, served by our build.
registry_build: runtime_archs now counts an architecture only when llama_model_mapping builds a model for it. Upstream's name-only gptj was being counted. FORK_ONLY matches serve's list.
registry/architectures.json: 269 → 321 HF architectures mapped. The gap check passes.

🤖 Generated with Claude Code

… names (llama.cpp #8) The pin adds four architectures upstream llama.cpp has no model for (gptj only a name), each checked against transformers FP32 before merging (teacher-forced top-1: OPT-125m 96/96, GPT-Neo-125m 96/96, CodeGen-350M 96/96, GPT-J tiny 89/89), and registers sibling spellings of HF classes the runtime already has. - serve: fork_only_arch now covers opt, codegen, gptneo and gptj next to zaya, so --device vulkan runs them on our build (checked: OPT-125m and GPT-Neo-125m answer through 1bit serve --device vulkan with the upstream Vulkan server present). - registry_build: runtime_archs counts an architecture only when llama_model_mapping builds a model for it (upstream names gptj but has no class); FORK_ONLY matches serve. - registry/architectures.json: 269 -> 321 HF architectures mapped. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

context7 · 2026-09-26T14:16:41Z

Docs7 for 1bit-monster/engine

Result	Status	Action
Deployment	➖ Not used	—
Content review	➖ Did not run. This site has no agent runs available this month. Wait for the monthly reset or check your Docs7 plan.	—

Commit 3a8290f

github-actions · 2026-09-26T14:17:47Z

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 8 - Partially compliant Compliant requirements: Pin llama.cpp to commit 83e1c41 Add support for OPT, CodeGen, GPT-Neo, and GPT-J architectures Update registry_build.py to reflect new architectures Update architectures.json to include new architectures Ensure fork_only_arch function in serve.cpp includes new architectures Non-compliant requirements: None Requires further human verification: None
⏱️ Estimated effort to review: 3 🔵🔵🔵⚪⚪
🧪 No relevant tests
🔒 No security concerns identified
⚡ Recommended focus areas for review Architecture List Synchronization The `fork_only_arch` function in `app/serve.cpp` is updated to include new architectures (opt, codegen, gptneo, gptj) to match the `FORK_ONLY` set in `tools/registry_build.py`. This change ensures that the architecture list is synchronized across the codebase. However, it's important to verify that all new architectures are correctly handled by the runtime and that no unintended architectures are included. bool fork_only_arch(const std::string& gguf) { // Zyphra ZAYA1 (docs/vulkan.md); OPT, CodeGen, GPT-Neo and GPT-J (llama.cpp #8: upstream has // no model for them, gptj only a name). Keep in step with FORK_ONLY in tools/registry_build.py. static const char* const archs[] = {"zaya", "opt", "codegen", "gptneo", "gptj"}; const std::string arch = gguf_architecture(gguf); for (const char* a : archs) if (arch == a) return true; return false; } Runtime Architecture Detection Logic The `runtime_archs` function in `tools/registry_build.py` has been updated to detect architectures that the runtime can build a model for by parsing `llama-arch.cpp` and `llama-model.cpp`. This change ensures that the registry accurately reflects which architectures are supported by the runtime. It's important to verify that the parsing logic correctly identifies all architectures that are actually supported. def runtime_archs(tree): """GGUF architecture names the runtime can build a model for: named in LLM_ARCH_NAMES and given a class in llama_model_mapping (a name alone, as upstream has for gptj, loads nothing).""" src = open(os.path.join(tree, "src/llama-arch.cpp")).read() body = src[src.index("LLM_ARCH_NAMES"):] body = body[:body.index("};")] names = dict(re.findall(r"\{\s(LLM_ARCH_[A-Z0-9_]+)\s,\s\"([^\"]+)\"\s\}", body)) model = open(os.path.join(tree, "src/llama-model.cpp")).read() start = model.index("llama_model_mapping(") mapping = model[start:model.index("\n}\n", start)] built = set(re.findall(r"case\s+(LLM_ARCH_[A-Z0-9_]+)\s*:", mapping)) return {n for e, n in names.items() if e in built} - {"clip"}

… registry counts current (#162) * Site: this week's milestones post; OPT/GPT-Neo/CodeGen/GPT-J in docs; registry counts current - blog: Zyphra's whole family (74B, legacy base models, ZAYA1-VL, Zamba*), four older architectures, MoE experts from the drive, deterministic HRX decode, ONNX on the GPU; every number from a committed doc. - docs/vulkan.md: the four older architectures from our llama.cpp (#138), with their FP32 agreement. - docs/registry.md: counts at the current pins (323 mapped; was 265), and what mapped means. - registry: regenerated for ZINC 29bc350 (#132); counts unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * registry: regenerate for llama.cpp-vulkan 546800c (#161) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: bong-water-water-bong <bong-water-water-bong@1bit.gg> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

github-actions Bot added the Review effort 3/5 label Sep 26, 2026

bong-water-water-bong merged commit 4b707c0 into main Sep 26, 2026
3 checks passed

bong-water-water-bong deleted the pin-fork-archs branch September 26, 2026 14:18

bong-water-water-bong mentioned this pull request Sep 26, 2026

Site: this week's milestones post; OPT/GPT-Neo/CodeGen/GPT-J in docs; registry counts current #162

Merged

bong-water-water-bong mentioned this pull request Sep 27, 2026

docs: census figures of 2026-09-27 (94.88% mapped, 64.33% checked) #172

Merged

bong-water-water-bong mentioned this pull request Oct 30, 2026

Bump HRX: llama.cpp d8d2f5b39f22 + hrx-system 82804197bcbe #659

Merged

Navigation Menu

Repository navigation

Pin llama.cpp 83e1c41: OPT, CodeGen, GPT-Neo, GPT-J and 48 sibling HF names - #138

Pin llama.cpp 83e1c41: OPT, CodeGen, GPT-Neo, GPT-J and 48 sibling HF names#138
bong-water-water-bong merged 1 commit into
mainfrom
pin-fork-archs

bong-water-water-bong commented Sep 26, 2026

Uh oh!

context7 Bot commented Sep 26, 2026

Uh oh!

github-actions Bot commented Sep 26, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

bong-water-water-bong commented Sep 26, 2026

Uh oh!

context7 Bot commented Sep 26, 2026

Uh oh!

github-actions Bot commented Sep 26, 2026

PR Reviewer Guide 🔍

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant