pytorch-labs float8_experimental issues

pytorch-labs / float8_experimental

This repository contains the experimental PyTorch native float8 training UX

BSD 3-Clause "New" or "Revised" License

189 stars 18 forks source link

issues

Newest

Newest Most commented Recently updated Oldest Least commented Least recently updated

fix nits from deletion of Float8DynamicLinear

#308 vkuzo opened 2 hours ago
0
unify linear test cases

#307 vkuzo opened 7 hours ago
0
static scaling support for training

#306 vkuzo closed 2 hours ago
1
Add rowwise scaling to Float8Inference module

#305 drisspg opened 5 days ago
0
delete Float8DynamicLinear

#304 vkuzo closed 3 days ago
2
Add utility for filtering out skipped tests in large cross-product groups

#303 drisspg opened 5 days ago
0
Add sanity checks to dtensor tests

#302 drisspg closed 5 days ago
1
Thread the scaling type argument throughout fp8

#301 drisspg opened 5 days ago
0
[9/x]: make dynamic scaling default in Float8Linear

#300 vkuzo closed 5 days ago
2
[8/x] make single linear profiling script work with Float8 scaling type

#299 vkuzo closed 5 days ago
2
[7/x] make profiling script support Float8Linear dynamic scaling

#298 vkuzo closed 5 days ago
2
[6/x] switch inference tests to use Float8Linear

#297 vkuzo closed 5 days ago
2
[5/x] make FSDP2 with float8 all-gather work for Float8Linear

#296 vkuzo closed 5 days ago
2
Adds a test comparing the output of torch.compile and export

#295 drisspg opened 1 week ago
0
[4/x] add tests for DTensor TP/SP + Float8Linear

#294 vkuzo closed 5 days ago
2
[3/x]: simplify FSDP1 test and add coverage for dynamic scaling

#293 vkuzo closed 5 days ago
2
Float8Tensor.to_original_precision() returns wrong dtype

#292 ani300 opened 1 week ago
0
[2/x]: fix numerics integration test and test delayed vs dynamic

#291 vkuzo closed 5 days ago
2
[1/x]: Make Float8Linear support dynamic scaling

#290 vkuzo closed 5 days ago
2
make testing better on amd

#289 drisspg opened 2 weeks ago
0
Add a Float8LinearInference module to support static, dynamic, and wo quant

#287 drisspg closed 1 week ago
5
[ROCm] Unskip passing torch.compile test

#286 alugorey closed 2 weeks ago
2
Add more compile compatibility for Float8Tensor ops

#285 ani300 closed 1 week ago
10
Updates with new scaled-mm api

#284 drisspg closed 2 weeks ago
4
Adding Float8 Linear variants supporting inference-only with lower overhead

#283 cyang49 opened 3 weeks ago
1
add norm_ffn_norm to profile script

#282 vkuzo closed 1 week ago
2
QOL improvements to benchmarks/profile_linear_float8.py

#281 vkuzo closed 1 week ago
2
Docs should say what's the smallest model users will see a benefit for

#280 msaroufim opened 3 weeks ago
1
Expected trailing dimension of mat1 to be divisible by 16 but got mat1 shape

#279 msaroufim opened 3 weeks ago
3
QOL improvements to linear benchmarking script

#278 vkuzo closed 3 weeks ago
4
delayed scaling: stop syncing weight amax values across ranks

#277 vkuzo closed 3 weeks ago
2
delayed scaling: delete Float8LinearMixin

#276 vkuzo closed 3 weeks ago
2
add PrepareFloat8ModuleInput for sequence parallel

#275 wanchaol closed 3 weeks ago
3
[QST] Dynamic Scaling

#274 jeromeku opened 1 month ago
2
QOL improvements to linear benchmarking script

#273 vkuzo closed 3 weeks ago
1
delayed scaling: stop syncing weight amax values across ranks

#272 vkuzo closed 3 weeks ago
1
delayed scaling: delete Float8LinearMixin

#271 vkuzo closed 3 weeks ago
3
[not for land] enumerate breakages with module hooks + compile

#270 vkuzo opened 1 month ago
1
[not for land] testing ghstack 2

#269 vkuzo closed 1 month ago
0
[not for land] testing ghstack

#268 vkuzo closed 1 month ago
0
delayed scaling safety logic currently doesn't work with activation checkpointing

#267 vkuzo opened 1 month ago
0
[FSDP2] pre-compute amax after optimizer.step for dynamic scaling

#266 weifengpy opened 1 month ago
1
[FSDP2] set vocab_size=32 to avoid must be divisible by 16 error

#265 weifengpy closed 1 month ago
2
[FSDP2] set `vocab_size=32` to avoid `must be divisible by 16` error

#264 weifengpy closed 1 month ago
4
enable float types in pytorch for non comptue comms

#263 drisspg closed 1 month ago
3
add wait_tensor() after all_gather in float8 to fix mem leak

#262 bdhirsh closed 1 month ago
2
add wait_tensor() after all_gather in float8 to fix mem leak

#261 bdhirsh closed 1 month ago
0
[not for land] standalone repro of memory leak on float8 + compile + …

#260 vkuzo opened 1 month ago
0
memory alignment issue in torch.compile mode

#259 czmrand opened 1 month ago
0
[wip] make all 3 gemms in float8 linear configurable

#258 vkuzo opened 1 month ago
0