An open API service for providing issue and pull request metadata for open source projects.

GitHub / pytorch-labs/float8_experimental issues and pull requests

#353 - Update README.md

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#352 - [wip] add axiswise granularity to Float8Tensor

Pull Request - State: open - Opened by vkuzo about 1 year ago
Labels: CLA Signed

#351 - [wip] add axiswise granularity to Float8Tensor

Pull Request - State: closed - Opened by vkuzo about 1 year ago
Labels: CLA Signed

#350 - [6/x] clean up casting: rename delayed and dynamic casting functions

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#349 - [5/x] clean up casting: cast_to_float8_e4m3_dynamic -> cast_to_float8_dynamic

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#348 - [4/x] clean up casting: ToFloat8ConstrFunc -> hp_tensor_and_scale_to_float8

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#347 - [3/x] clean up casting functions: delete to_fp8_no_autograd

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 3 comments
Labels: CLA Signed, Merged

#346 - [2/x] clean up casting functions: delayed scaling

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 4 comments
Labels: CLA Signed, Merged

#345 - [1/x] clean up casting functions

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 4 comments
Labels: CLA Signed, Merged

#344 - bring back torch.autograd.Function for float8 matmul

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 4 comments
Labels: CLA Signed, Merged

#343 - [2/x] clean up casting functions: delayed scaling

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 1 comment
Labels: CLA Signed

#342 - [1/x] clean up casting functions

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 1 comment
Labels: CLA Signed

#341 - bring back torch.autograd.Function for float8 matmul

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed

#340 - [2/x] clean up casting functions: delayed scaling

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 1 comment
Labels: CLA Signed

#339 - [1/x] clean up casting functions

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 1 comment
Labels: CLA Signed

#338 - [wip] add scaling granularity

Pull Request - State: open - Opened by vkuzo about 1 year ago
Labels: CLA Signed

#337 - [bc-breaking] rename TensorScalingType->ScalingType, Float8TensorCastConfig->CastConfig

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#336 - bring back torch.autograd.Function for float8 matmul

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed

#335 - rename all variables to use input/weight/grad_output notation

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#334 - add per-gemm config to `Float8LinearConfig`

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#333 - [bc-breaking] rename `DelayedScalingRecipe` to `DelayedScalingConfig`

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#332 - [bc-breaking] rename `config.enable_fsdp_fp8_all_gather` to use `float8`

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#331 - Reduced CPU overhead in `precompute_float8_dynamic_scale_for_fsdp`

Pull Request - State: closed - Opened by awgu about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#330 - [bc-breaking] clarify public API of float8_experimental

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#329 - [bc-breaking] rename top level UX to `convert_to_float8_training`

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#328 - [bc-breaking] switch argument order to `module_filter_fn`

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#327 - [bc-breaking] move `emulate` flag to `Float8LinearConfig`

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#326 - [DISCUSSION] fix float8 all-gather in FSDP2 + TP: DTensor(WeightWithDynamicFloat8CastTensor)

Pull Request - State: open - Opened by weifengpy about 1 year ago
Labels: CLA Signed

#325 - [bc-breaking] move tensor scaling configuration to Float8LinearConfig

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#324 - [bc breaking] move most previously-global configs to Float8LinearConfig

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#323 - [bc breaking] change x, w, dL_dY variable names to input, weight, grad_output

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#322 - [bc breaking] unify filtering functions

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#321 - add unit tests for FSDP2 + torch.compile(transformer block)

Pull Request - State: closed - Opened by weifengpy about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#320 - fixes to matmul and linear benchmarks

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#319 - fix README.md description of `swap_linear_with_float8_linear`

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#318 - Any quality benchmarks?

Issue - State: closed - Opened by tsengalb99 about 1 year ago - 1 comment

#317 - update readme

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#316 - bring back torch.autograd.Function

Pull Request - State: open - Opened by vkuzo about 1 year ago
Labels: CLA Signed

#315 - make all 3 gemms in Float8Linear support configurability, not user facing

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#314 - [RFC] Float8 Inference

Issue - State: closed - Opened by drisspg about 1 year ago - 8 comments

#313 - fix linter error in CI

Pull Request - State: closed - Opened by weifengpy about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#312 - support delayed scaling of weight in float8 all-gather

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 4 comments
Labels: CLA Signed, Merged

#311 - delete `swap_linear_with_dynamic` from fsdp2 eager test case

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#310 - move WeightWithDynamicFloat8CastTensor to fsdp_utils.py

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#309 - one more delayed -> dynamic default update

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#308 - fix nits from deletion of Float8DynamicLinear

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#307 - unify linear test cases

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#306 - static scaling support for training

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 1 comment
Labels: CLA Signed

#305 - Add rowwise scaling to Float8Inference module

Pull Request - State: open - Opened by drisspg about 1 year ago
Labels: CLA Signed

#304 - delete Float8DynamicLinear

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#303 - Add utility for filtering out skipped tests in large cross-product groups

Pull Request - State: open - Opened by drisspg about 1 year ago
Labels: CLA Signed

#302 - Add sanity checks to dtensor tests

Pull Request - State: closed - Opened by drisspg about 1 year ago - 1 comment
Labels: CLA Signed

#301 - Thread the scaling type argument throughout fp8

Pull Request - State: open - Opened by drisspg about 1 year ago
Labels: CLA Signed

#300 - [9/x]: make dynamic scaling default in Float8Linear

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#299 - [8/x] make single linear profiling script work with Float8 scaling type

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#298 - [7/x] make profiling script support Float8Linear dynamic scaling

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#297 - [6/x] switch inference tests to use Float8Linear

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#296 - [5/x] make FSDP2 with float8 all-gather work for Float8Linear

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#295 - Adds a test comparing the output of torch.compile and export

Pull Request - State: open - Opened by drisspg about 1 year ago
Labels: CLA Signed

#294 - [4/x] add tests for DTensor TP/SP + Float8Linear

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#293 - [3/x]: simplify FSDP1 test and add coverage for dynamic scaling

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#292 - Float8Tensor.to_original_precision() returns wrong dtype

Issue - State: closed - Opened by ani300 about 1 year ago - 1 comment

#291 - [2/x]: fix numerics integration test and test delayed vs dynamic

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#290 - [1/x]: Make Float8Linear support dynamic scaling

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#287 - Add a Float8LinearInference module to support static, dynamic, and wo quant

Pull Request - State: closed - Opened by drisspg about 1 year ago - 5 comments
Labels: CLA Signed, Merged

#283 - Adding Float8 Linear variants supporting inference-only with lower overhead

Pull Request - State: closed - Opened by cyang49 about 1 year ago - 2 comments
Labels: CLA Signed

#282 - add norm_ffn_norm to profile script

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#281 - QOL improvements to benchmarks/profile_linear_float8.py

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 2 comments
Labels: CLA Signed, Merged

#280 - Docs should say what's the smallest model users will see a benefit for

Issue - State: closed - Opened by msaroufim about 1 year ago - 2 comments
Labels: documentation

#279 - Expected trailing dimension of mat1 to be divisible by 16 but got mat1 shape

Issue - State: closed - Opened by msaroufim about 1 year ago - 4 comments
Labels: documentation

#274 - [QST] Dynamic Scaling

Issue - State: closed - Opened by jeromeku about 1 year ago - 3 comments

#267 - delayed scaling safety logic currently doesn't work with activation checkpointing

Issue - State: closed - Opened by vkuzo about 1 year ago - 1 comment

#266 - [FSDP2] precompute scale after optimizer.step for dynamic scaling

Pull Request - State: closed - Opened by weifengpy about 1 year ago - 7 comments
Labels: CLA Signed, Merged

#259 - memory alignment issue in torch.compile mode

Issue - State: closed - Opened by czmrand about 1 year ago - 1 comment

#258 - [wip] make all 3 gemms in float8 linear configurable

Pull Request - State: closed - Opened by vkuzo about 1 year ago - 1 comment
Labels: CLA Signed

#257 - Float8Linear does not support autocast

Issue - State: closed - Opened by yitzhaklevi about 1 year ago - 2 comments

#254 - Better default for DelayedScalingRecipe.history_len

Issue - State: closed - Opened by vkuzo about 1 year ago

#252 - Added func for precomputing dynamic fp8 casts

Pull Request - State: closed - Opened by awgu over 1 year ago
Labels: CLA Signed

#249 - investigate cuda graphs + dynamic scaling leading to memory fragmentations

Issue - State: closed - Opened by vkuzo over 1 year ago - 1 comment

#246 - add configuration of precision for all 3 gemms

Issue - State: closed - Opened by vkuzo over 1 year ago - 1 comment

#245 - add option to use fast accumulation in the float8 matmul

Issue - State: closed - Opened by vkuzo over 1 year ago - 2 comments

#244 - clean up and document skipping layers logic

Issue - State: closed - Opened by vkuzo over 1 year ago - 1 comment

#243 - addmm implemented incorrectly

Issue - State: closed - Opened by lw over 1 year ago - 1 comment

#242 - Investigate Sam test tolerances

Issue - State: closed - Opened by drisspg over 1 year ago - 1 comment

#238 - torch.inference_mode switches`aten.linear.default, this is not supported`

Issue - State: closed - Opened by michaelfeil over 1 year ago - 4 comments

#237 - Graph breaks in Float8Linear code

Issue - State: closed - Opened by y-sq over 1 year ago - 1 comment

#226 - Issue with running isnan and _is_any_true with Float8Tensor

Issue - State: closed - Opened by snarayan21 over 1 year ago - 5 comments

#223 - Use Activation Hooks failing with AotAutograd for dynamic linear

Issue - State: closed - Opened by drisspg over 1 year ago - 4 comments
Labels: Compile

#207 - torch.compile() not working with Float8Linear

Issue - State: closed - Opened by yitzhaklevi over 1 year ago - 2 comments

#202 - Float8Tensor + DTensor composability

Issue - State: closed - Opened by wanchaol over 1 year ago - 1 comment

#201 - Float8 + Per-Parameter FSDP

Issue - State: closed - Opened by drisspg over 1 year ago - 1 comment
Labels: Tracking

#199 - Hugging Face transformers integration

Issue - State: closed - Opened by younesbelkada over 1 year ago - 1 comment

#195 - Top Level Torch Compile Issue Tracker

Issue - State: closed - Opened by drisspg over 1 year ago - 1 comment
Labels: Tracking

#187 - upcoming feature tracker

Issue - State: closed - Opened by vkuzo over 1 year ago - 3 comments
Labels: Tracking

#185 - Enable fp8_weight recomputation during backwards pass

Issue - State: closed - Opened by drisspg over 1 year ago - 2 comments

#174 - [not for land, yet] example of FSDP with dynamic float8 linear

Pull Request - State: closed - Opened by vkuzo over 1 year ago
Labels: CLA Signed

#156 - Revisit approach to keeping amax/scale buffers in fp32

Issue - State: closed - Opened by awgu over 1 year ago - 1 comment

#119 - Speedup sync_float8_amax_and_scale_history

Issue - State: closed - Opened by drisspg almost 2 years ago - 3 comments
Labels: Perf

#111 - Support for Fused Attention + FP8

Issue - State: closed - Opened by drisspg almost 2 years ago - 1 comment
Labels: Perf

#100 - make benchmarking script output nicer

Pull Request - State: closed - Opened by vkuzo almost 2 years ago
Labels: CLA Signed