ahrefs ocannl issues - Githubissues

ahrefs / ocannl

OCANNL: OCaml Compiles Algorithms for Neural Networks Learning

BSD 2-Clause "Simplified" License

67 stars 2 forks source link

issues

Newest

Newest Most commented Recently updated Oldest Least commented Least recently updated

Consider measuring the overhead of creating (recording) events

#292 lukstafi opened 2 weeks ago
0
Audit manually specifying sharing in a tensor node memory mode

#291 lukstafi opened 2 weeks ago
0
The use of `shared_merge_buffer` should depend on the tensor node's `memory_mode` (on all backends)

#290 lukstafi opened 2 weeks ago
0
Make memory reporting for CUDA more meaningful

#289 lukstafi opened 1 month ago
0
Consider introducing proper static verification of merge buffer nodes, but "in the right direction"

#288 lukstafi opened 1 month ago
0
Rename (virtual) `device` -> `stream`, and `physical_device` -> `device`

#287 lukstafi closed 1 month ago
3
Incorporate cross-stream-sharing into the memory model in `Tnode`

#286 lukstafi closed 1 month ago
0
Rename `writing_a_backend.md` to `anatomy_of_a_backend.md` and update to reflect the new design of the multi-stream CUDA backend

#285 lukstafi closed 3 weeks ago
0
Support half precision in the GCCJIT backend

#284 lukstafi opened 2 months ago
0
Fix too-tight patterns in `ppx_op.ml` so that examples from the documentation work

#283 lukstafi closed 2 months ago
0
ppx_op: When an argument with the label `~config` is applied to an expression, lift the application out of functions

#282 lukstafi closed 2 months ago
1
When in a label the same identifier repeats consecutively (e.g. `ident` 3 times), let `debug_name` collapse it (e.g. `ident3`)

#281 lukstafi opened 3 months ago
0
ppx_op: Include `config.label` in the label of the primary (returned) tensor, when available, and not incorporating a label from a function argument

#280 lukstafi opened 3 months ago
1
Remove the `ident_label` functionality from `ppx_cd`, as it almost never has a "primary" created tensor

#279 lukstafi closed 2 months ago
0
Example training loop using DisTrO low-communication distributed data parallelism

#278 lukstafi opened 3 months ago
0
Anything we can learn from `krnl` and `autograph`?

#277 lukstafi opened 4 months ago
0
Simplify `Tnode.task.work` once ppx_minidebug supports domain-local debug runtimes

#276 lukstafi closed 3 months ago
1
Replicate Andrej Karpathy's "LLM101n: Let's build a Storyteller"

#275 lukstafi opened 4 months ago
0
Make `unsafe_cleanup` safe, get rid of `unsafe_shutdown`, initialize devices on-demand

#274 lukstafi closed 4 months ago
1
Get rid of `Utils.settings.with_debug_level`, instead unify with ppx_minidebug debug levels

#273 lukstafi closed 3 months ago
2
Require that tensor nodes in the graph of a tensor are either embedded, or already part of the parent context

#272 lukstafi closed 4 months ago
1
Support quantization for optimizers: low-bit optimizers

#271 lukstafi opened 4 months ago
0
Any lessons from Imbue for training-in-the-large?

#270 lukstafi opened 4 months ago
1
CI failing on MacOS for unknown reason

#269 lukstafi closed 2 months ago
3
Refactor backends to two layers of sharing: C-syntax ("c", "cuda", "metal") and CUDA-like ("cuda", "metal"?)

#268 lukstafi closed 4 months ago
1
Take a look at Tiramisu Polyhedral Compiler

#267 lukstafi opened 6 months ago
0
Consider implementing a Cranelift backend (CPU)

#266 lukstafi opened 6 months ago
0
Study Candle -- a minimalistic Rust framework

#265 lukstafi opened 6 months ago
0
Implement virtual CUDA devices via streams

#264 lukstafi closed 4 months ago
1
Consider implementing Lean Attention (Flash Attention + softmax-as-reduce)

#263 lukstafi opened 6 months ago
0
Implement "parameter punning" for the code notation `%cd`

#262 lukstafi closed 3 months ago
1
Superoptimizers for tensor programs

#261 lukstafi opened 6 months ago
0
Implement an xLSTM LLM

#260 lukstafi opened 6 months ago
1
Consider reducing the dependencies on JaneStreet libraries; especially, break-up `ppx_jane`

#259 lukstafi closed 1 month ago
2
`bin/einsum_trivia.ml`: High-level code is missing from `c_fwd.cd`

#258 lukstafi closed 7 months ago
0
Resolves #256 by deleting all occurrences of Shape.backprop_ith_arg in the project

#257 mikhailazaryan closed 7 months ago
2
No one is using `Shape.backprop_ith_arg`. Why? Remove it.

#256 lukstafi closed 7 months ago
5
Audit and/or more extensively test dimension label checking and inference

#255 lukstafi opened 7 months ago
0
Introduce a division operator that raises a shape error when division is with non-zero remainder

#254 lukstafi opened 7 months ago
0
Study and incorporate Andrej Karpathy's `llm.c` lessons

#253 lukstafi opened 7 months ago
3
Special-case `let%cd _ = ...` to not modify (forward, backward) roots

#252 lukstafi closed 7 months ago
1
Introduce a `Reverse_eq` constraint: r1 is equal to the reverse of r2

#251 lukstafi closed 7 months ago
0
Sanitizer mode for synchronizing tnodes across devices

#250 lukstafi closed 7 months ago
0
Both `dim_constraint` and `row_constraint` (former `dims_constraint`) should be stored in entries

#249 lukstafi closed 7 months ago
1
Move the `%cd` syntax to the arrayjit library / package

#248 lukstafi closed 4 months ago
2
Consider adding cycle detection for row variables

#247 lukstafi opened 8 months ago
0
Fix performance of `bin/micrograd_basic` -- very simple scalar computation takes 7 seconds to compile

#246 lukstafi closed 8 months ago
1
Fix measuring memory consumption in benchmarks

#245 lukstafi closed 1 month ago
0
Do a deep dive into Luminal

#244 lukstafi opened 8 months ago
0
Figure out a way to only generate one file per each of `.ll`, `.cd` for sources routed through `prejit_batch` or at least `merge_batch`

#243 lukstafi closed 8 months ago
1