v0.14.9

Flux v0.14.9

Diff since v0.14.8

Merged pull requests:

• Restore type stability of conv_transpose_dims (#2365) (@ToucheSir)
• Hotfix for new OneElement on GPU (#2368) (@ToucheSir)

v0.14.8

Flux v0.14.8

Diff since v0.14.7

Merged pull requests:

• Use stable API for AMDGPU RNG conversion (#2360) (@ToucheSir)
• Non-diff shape handling in norm layers (#2363) (@ToucheSir)

Closed issues:

• 0.2 tutorial at the top of google for "flux mnist" (#2163)
• Illegal Memory Access Error During Gradient Calculation of predefined losses on GPU RTX 4050 (#2361)
• Unnecessarily using shared GPU memory (#2364)

v0.14.7

Flux v0.14.7

Diff since v0.14.6

Merged pull requests:

• Bump actions/checkout from 4.0.0 to 4.1.0 (#2340) (@dependabot[bot])
• Use new public feature (#2342) (@mcabbott)
• Bump thollander/actions-comment-pull-request from 2.4.2 to 2.4.3 (#2347) (@dependabot[bot])
• Bump actions/checkout from 4.1.0 to 4.1.1 (#2348) (@dependabot[bot])
• Fix test that is not broken anymore (#2349) (@devmotion)
• CompatHelper: add new compat entry for Statistics at version 1, (keep existing compat) (#2351) (@github-actions[bot])
• Fix typo in docs/src/tutorials/2021-01-26-mlp.md (#2353) (@poludmik)
• Fixing typos in documentation (#2355) (@poludmik)
• Bump AMDGPU.jl compat to 0.7 (#2356) (@pxl-th)
• Bump AMDGPU compat to 0.8 (#2359) (@pxl-th)

Closed issues:

• Android/iOS support (#2357)

v0.14.6

Flux v0.14.6

Diff since v0.14.5

Merged pull requests:

• Adding tooling for JuliaFormatter. (#2323) (@codetalker7)
• Fix typo (#2329) (@christiangnrd)
• Bump AMDGPU compat to 0.6 (#2332) (@pxl-th)
• Bump dorny/paths-filter from 2.9.1 to 2.11.1 (#2333) (@dependabot[bot])
• Bump actions/checkout from 2.2.0 to 4.0.0 (#2334) (@dependabot[bot])
• update CUDA compat (#2338) (@CarloLucibello)

Closed issues:

• move the codebase to 4 spaces indentation (#2298)
• No cudnn implementation of Conv((1,) N=>M) (#2322)
• can't use masks in multi-head-attention layer (#2336)
• Adapt saving & loading example to CuArrays (#2337)
• Segmentation fault when doing a forward pass with a model saved with BSON (#2339)

v0.14.5

Flux v0.14.5

Diff since v0.14.4

Merged pull requests:

• Bump actions/checkout from 1.0.0 to 3.6.0 (#2324) (@dependabot[bot])
• Bump actions/checkout from 3.6.0 to 4.0.0 (#2326) (@dependabot[bot])
• rename "AMD" backend to "AMDGPU" (#2328) (@CarloLucibello)

v0.14.4

Flux v0.14.4

Diff since v0.14.3

Merged pull requests:

• allow get_device("Metal") and informative error messages (#2319) (@CarloLucibello)

v0.14.3

Flux v0.14.3

Diff since v0.14.2

Closed issues:

• No error from negative learning rates (#1982)
• Implement data movement across GPU devices. (#2302)
• train! using Metal and stateful optimizers fails (#2310)
• Warning: sort(d::Dict; args...) is deprecated, use sort!(OrderedDict(d); args...) instead. (#2312)
• Does withgradient have lower precicision than simply calling the function? (#2315)
• Warning: sort(d::Dict; args...) is deprecated, use sort!(OrderedDict(d); args...) instead. (#2320)

Merged pull requests:

• Bump thollander/actions-comment-pull-request from 2.4.0 to 2.4.2 (#2307) (@dependabot[bot])
• Implement interface for data transfer across GPU devices. (#2308) (@codetalker7)
• Removing deprecated method call in GPU_BACKEND_ORDER. (#2314) (@codetalker7)
• Added entry for RobustNeuralNetworks.jl in ecosystem.md (#2317) (@nic-barbara)
• Allow Optimisers.jl v0.3 (#2318) (@CarloLucibello)

v0.14.2

Flux v0.14.2

Diff since v0.14.1

Closed issues:

• Mixed precision training. (#543)
• have buildkite run GPU tests only (#2271)
• Allow old silent behavior for gpu (#2293)
• @autosize macro is not working (#2296)
• Why Flux is Significantly Slower than Pytorch? (#2300)
• Huber Loss Fails with Metal GPU (#2305)

Merged pull requests:

• Adding device objects for selecting GPU backends (and defaulting to CPU if none exists). (#2297) (@codetalker7)
• Run only GPU tests on buildkite. (#2301) (@codetalker7)
• Avoid broadcast-related type instabilities with huber_loss (#2306) (@jeremiahpslewis)

v0.14.1

Flux v0.14.1

Diff since v0.14.0

Closed issues:

• gpu() returns nothing (#2294)

Merged pull requests:

• Update NEWS.md for 0.14 release (#2288) (@mcabbott)
• Update description of trainable in "advanced.md" (#2289) (@mcabbott)
• fix: make gpu(x) return unmodified x when GPU backends aren't loaded (#2295) (@IanButterworth)

v0.14.0

Flux v0.14.0

Diff since v0.13.17

Flux now requires Julia 1.9, to take advantage of package extensions. CUDA is no longer loaded automatically, to speed up loading when not using a GPU / using a non-Nvidia one.

Previously deprecated functions removed are:

• Flux.stop and Flux.skip, in favour of break / continue
• The macro @epochs, in favour of a for loop
• Flux.zeros and Flux.ones, in favour of zeros32 and ones32.

Closed issues:

• Choosing model serialization format(s) for cross-framework support (like HuggingFace) (#1907)
• Very slow "using Flux" (#1961)
• SSIM loss (#2165)
• buildkite failure on julia v1.6 (#2214)
• Sysimage compilaiton failed (#2242)
• PackageCompiler fails with Flux on embedded ARM/no GPU (#2262)
• how to make CUDA functionalities an extension (#2265)
• Recurrent layers can't be applied to views of OneHotArrays (#2279)
• Question about using loop in loss function (#2280)

Merged pull requests:

• add CUDA extension (#2268) (@CarloLucibello)
• fix doc of PairwiseFusion (#2281) (@ctarn)
• Update AMDGPU & Metal compat & add CI job (#2282) (@pxl-th)
• Cleanup for v0.14 release (#2283) (@CarloLucibello)
• tag v0.14 (#2284) (@CarloLucibello)
• Update training.md (#2286) (@dreivmeister)
• Fix typos in 0.14 docs (#2287) (@christiangnrd)