Project status

[ggerganov]

[NO LONGER UPDATED]

Below is a summary of the functionality provided by the llama.cpp project.

• The goal is to have a birds-eye-view of what works and what does not
• Collaborators are encouraged to add things to the list and update the status of existing things as needed
• The list should be simple without too much details about the specific problems - these belong to dedicated issues

Legend (feel free to update):

✅ - Working correctly
☁️ - Partially working
❌ - Failing
❓ - Status unknown (needs testing)
🔬 - Under investigation
🚧 - Currently in development

Feature	Executable	Status	Issues
Inference
Single-batch decoding	main, simple	✅
Parallel / batched decoding	batched	✅
Continuous batching	parallel	✅
Speculative sampling	speculative	✅
Tree-based speculative sampling	speculative	✅
Self-speculative sampling	speculative	🚧	#3565
Lookahead sampling	lookahead	✅
Infill	infill	✅
REST API	server	✅
Embeddings	embedding	✅
Grouped Query Attention CPU	main	✅
Grouped Query Attention CUDA	main	✅
Grouped Query Attention OpenCL	main	✅
Grouped Query Attention Metal	main	✅
Session load / save	main	✅
K-quants (256) CUDA	main	✅
K-quants (64) CUDA	main	✅
K-quants (256) Metal	main	✅
K-quants (64) Metal	main	☁️	#3276
Special tokens	main	✅
Grammar sampling	main, server	✅
Beam search	beam-search	❓	#3471 (comment)
LoRA	main	☁️	#3333 #3519
SPM tokenizer	test-tokenizer-0-llama	✅
BPE tokenizer	test-tokenizer-0-falcon	✅
Models
LLaMA v1	main	✅
LLaMA v2	main	✅
Falcon	main	✅
StarCoder	main	✅
Baichuan	main	✅
MPT	main	✅
Persimmon	main	✅
LLaVA	llava	✅
Refact	main	✅
Bloom	main	✅
StableLM-3b-4e1t	main	✅
Training
Finetuning CPU	finetune	✅
Finetuning Metal	finetune	🔬
Backends
CPU x64	ggml	✅
CPU Arm	ggml	✅
GPU CUDA	ggml-cuda	✅
GPU ROCm	ggml-cuda	✅
GPU Metal	ggml-metal	✅
GPU OpenCL	ggml-opencl	✅
GPU Vulkan	ggml-vulkan	🚧	#2059

[niansa]

What does the "☁️" mean?

[shibe2]

I don't know what the icon means, but current status of OpenCL back-end is: it works with supported models, but is buggy and perhaps, slower than it could be.

[ggerganov]

Yup, this was my impression from reading a few issues lately. If you think it's not the case, feel free to update it. I just haven't set up OpenCL in my environment and cannot do tests atm

[ScarletEmerald]

So "Parallel decoding" is done by batched and "Continuous batching" is done by parallel? Are these reversed?

[ggerganov]

Parallel decoding is also called "batched decoding" hence batched. The parallel example demonstrates a basic server that serves clients in parallel - it just happens to have the continuous batching feature as an option.

Naming things is hard :) Sorry if these are confusing

[slaren]

Should beam search be added here? I think it is broken atm, at least with CUDA.

[ggerganov]

Yes, it should be added. The list is far from complete

[Mihaiii]

Fwiw, for me beam search is broken even without CUDA in a sense that when I run the example, nothing happens (it just hangs for minutes at this line until I CTRL+C it).

If it's an unknown problem, I'll open an issue (tbh, it's strange that nobody mentioned it before so maybe I'm doing something wrong).

Update: when it hangs on the above mentioned line, I have 0 hard page fauls/sec.

[slaren]

With CUDA it works for a while, but then it starts generating gibberish. I think that the calls to llama_decode are failing and it is not catching it. It's probably missing some KV cache management after the batched decoding change.

[ggerganov]

The beam search functionality should be moved out from the library and implemented as a standalone example.

[shibe2]

What would be criteria for considering OpenCL back-end working correctly? I've fixed all known bugs in ggml-opencl.cpp and now working on refactoring like #3669.

[ggerganov]

The criteria is that if it runs correctly on your machine, then it is ✅ until someone reports a problem that is reproducible - then it becomes ☁️ or ❌ depending on how broken the thing is

[shibe2]

Alright, turning the green light then!

[Yossef-Dawoad]

maybe you can ditch the icons for something Like Scoring Like (A+, A, A-, B, ...) this will make it obvious if something working fine but needs improvements has a score with A- and so on, maybe something like this :
[ A+ ] or [ A ] : working like charm
[ A- ] : Working correctly but needs improvement
[ B ] : Partially working
[ B- ] : Partially working with big issues to be resolved
[ C ] : Status unknown (needs testing)
[ D+ ] : Under investigation
[ D ] : Currently in development
[ F ] : Failing

maybe you should add a column for tier support for example, if a feature is tier 1 or 2, ... what do you think?

[pareek-ml]

Is there any further progress on Finetuning on metal gpu?