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

211 stars 20 forks source link

issues

Newest

Newest Most commented Recently updated Oldest Least commented Least recently updated

Update README.md

#353 vkuzo closed 3 months ago
2
[wip] add axiswise granularity to Float8Tensor

#352 vkuzo opened 3 months ago
0
[wip] add axiswise granularity to Float8Tensor

#351 vkuzo closed 3 months ago
0
[6/x] clean up casting: rename delayed and dynamic casting functions

#350 vkuzo closed 3 months ago
2
[5/x] clean up casting: cast_to_float8_e4m3_dynamic -> cast_to_float8_dynamic

#349 vkuzo closed 3 months ago
2
[4/x] clean up casting: ToFloat8ConstrFunc -> hp_tensor_and_scale_to_float8

#348 vkuzo closed 3 months ago
2
[3/x] clean up casting functions: delete to_fp8_no_autograd

#347 vkuzo closed 3 months ago
3
[2/x] clean up casting functions: delayed scaling

#346 vkuzo closed 3 months ago
4
[1/x] clean up casting functions

#345 vkuzo closed 3 months ago
4
bring back torch.autograd.Function for float8 matmul

#344 vkuzo closed 3 months ago
4
[2/x] clean up casting functions: delayed scaling

#343 vkuzo closed 3 months ago
1
[1/x] clean up casting functions

#342 vkuzo closed 3 months ago
1
bring back torch.autograd.Function for float8 matmul

#341 vkuzo closed 3 months ago
2
[2/x] clean up casting functions: delayed scaling

#340 vkuzo closed 3 months ago
1
[1/x] clean up casting functions

#339 vkuzo closed 3 months ago
1
[wip] add scaling granularity

#338 vkuzo opened 3 months ago
0
[bc-breaking] rename TensorScalingType->ScalingType, Float8TensorCastConfig->CastConfig

#337 vkuzo closed 3 months ago
2
bring back torch.autograd.Function for float8 matmul

#336 vkuzo closed 3 months ago
2
rename all variables to use input/weight/grad_output notation

#335 vkuzo closed 3 months ago
2
add per-gemm config to `Float8LinearConfig`

#334 vkuzo closed 3 months ago
2
[bc-breaking] rename `DelayedScalingRecipe` to `DelayedScalingConfig`

#333 vkuzo closed 3 months ago
2
[bc-breaking] rename `config.enable_fsdp_fp8_all_gather` to use `float8`

#332 vkuzo closed 3 months ago
2
Reduced CPU overhead in `precompute_float8_dynamic_scale_for_fsdp`

#331 awgu closed 3 months ago
2
[bc-breaking] clarify public API of float8_experimental

#330 vkuzo closed 3 months ago
2
[bc-breaking] rename top level UX to `convert_to_float8_training`

#329 vkuzo closed 3 months ago
2
[bc-breaking] switch argument order to `module_filter_fn`

#328 vkuzo closed 3 months ago
2
[bc-breaking] move `emulate` flag to `Float8LinearConfig`

#327 vkuzo closed 3 months ago
2
[DISCUSSION] fix float8 all-gather in FSDP2 + TP: DTensor(WeightWithDynamicFloat8CastTensor)

#326 weifengpy opened 3 months ago
0
[bc-breaking] move tensor scaling configuration to Float8LinearConfig

#325 vkuzo closed 3 months ago
2
[bc breaking] move most previously-global configs to Float8LinearConfig

#324 vkuzo closed 3 months ago
2
[bc breaking] change x, w, dL_dY variable names to input, weight, grad_output

#323 vkuzo closed 3 months ago
2
[bc breaking] unify filtering functions

#322 vkuzo closed 3 months ago
2
add unit tests for FSDP2 + torch.compile(transformer block)

#321 weifengpy closed 4 months ago
2
fixes to matmul and linear benchmarks

#320 vkuzo closed 4 months ago
2
fix README.md description of `swap_linear_with_float8_linear`

#319 vkuzo closed 4 months ago
2
Any quality benchmarks?

#318 tsengalb99 closed 3 months ago
1
update readme

#317 vkuzo closed 4 months ago
2
bring back torch.autograd.Function

#316 vkuzo opened 4 months ago
0
make all 3 gemms in Float8Linear support configurability, not user facing

#315 vkuzo closed 4 months ago
2
[RFC] Float8 Inference

#314 drisspg closed 3 months ago
8
fix linter error in CI

#313 weifengpy closed 4 months ago
2
support delayed scaling of weight in float8 all-gather

#312 vkuzo closed 4 months ago
4
delete `swap_linear_with_dynamic` from fsdp2 eager test case

#311 vkuzo closed 4 months ago
2
move WeightWithDynamicFloat8CastTensor to fsdp_utils.py

#310 vkuzo closed 4 months ago
2
one more delayed -> dynamic default update

#309 vkuzo closed 4 months ago
2
fix nits from deletion of Float8DynamicLinear

#308 vkuzo closed 4 months ago
2
unify linear test cases

#307 vkuzo closed 4 months ago
2
static scaling support for training

#306 vkuzo closed 4 months ago
1
Add rowwise scaling to Float8Inference module

#305 drisspg opened 4 months ago
0
delete Float8DynamicLinear

#304 vkuzo closed 4 months ago
2