Release notes

AMD publishes two families of Primus training Docker image:

Image	Backends	Documented in
`rocm/primus:<version>`	Megatron-LM, TorchTitan, Megatron Bridge	Megatron-LM, TorchTitan
`rocm/jax-training:maxtext-<version>`	MaxText (JAX)	JAX MaxText

The two families share a version number but are not built in lockstep — the same vNN.N tag can be weeks apart between families, and the MaxText family also ships patch releases (v26.3.1, v26.3.2) with no rocm/primus counterpart. Always read the section for the exact tag you are running.

This page is the single source of truth for image contents. Other pages link here instead of repeating version tables, so there is exactly one place to update per release. If you add a page that names an image tag, link to the relevant section below rather than restating the stack.

Every version below was read out of the published image itself. For v26.4 through v26.6 the values were additionally cross-checked against the release Dockerfiles in .github/workflows/docker-release/; earlier releases predate those files and are image-derived only. See Verifying the stack in an image to reproduce any table.

v26.6 (current)

`rocm/primus:v26.6`

Megatron-LM, TorchTitan, and Megatron Bridge backends.


Image ID	`4fcb3f210dc6`
Built	2026-08-25
Size	54.0 GB
Manifest	`f756f2279d8ab57b6549bcbd50d249755b69a407`
Dockerfile	`Dockerfile.primus-v26.6`

Software component	Version
ROCm	7.15.0 (`rocm-sdk` 7.15.0a20260727)
Python	3.12.3
PyTorch	2.12.0+rocm7.15.0a20260727
Transformer Engine	2.17.0+rocm7.15.0a20260727.e028a6c
Flash Attention	2.8.1
hipBLASLt	1.4.1-bbb68174
Triton	3.8.0+git4cff872c.rocm7.15.0a20260727
RCCL	2.30.4
torchvision	0.27.0+rocm7.15.0a20260727
torchaudio	2.11.0+rocm7.15.0a20260728
APEX	1.15.0a0+rocm10.1.0a20260822
AITER	0.1.14.post1
Primus-Turbo	0.4.1.dev26
torchao	0.15.0+gite9c7bead9
FBGEMM	2026.8.25
mamba-ssm / causal-conv1d / grouped_gemm	2.3.1 / 1.5.0.post8 / 1.1.4
transformers / datasets	5.5.0 / 3.6.0
NumPy	2.5.2

`rocm/jax-training:maxtext-v26.6`

MaxText (JAX) backend. The published tag is the MaxDiffusion-combined image (/workspace/maxdiffusion at 68e06965); MaxText remains at /workspace/maxtext.


Image ID	`a71d8dbb045e`
Built	2026-08-28
Size	42.6 GB
Manifest	`31a3a21d0a37cbd0b5de0342535f65d557ff77ba`
Dockerfile	`Dockerfile.jax-v26.6`

Software component	Version
ROCm	7.14.0
Python	3.12.3
JAX / jaxlib	0.11.0
jax-rocm7-pjrt / jax-rocm7-plugin	0.11.0.post1
Transformer Engine	2.17.0+rocm7.14.0.50a84ad
hipBLASLt	1.4.1-cd957402
RCCL	2.30.4 (built from rocm-systems `9e5e4084`)
Flax	0.12.8
TensorFlow	2.21.0 (CPU-only, rebuilt from the ROCm fork)
Optax / Orbax / Grain / tensorstore	0.2.8 / 0.12.4 / 0.2.18 / 0.1.85
MaxText	`b47d74bf` (`release/v26.6`)
transformers / datasets	4.57.3 / 4.8.5
NumPy	2.5.2

Note: transformers is 4.57.3 because the published image includes the MaxDiffusion stage, which pins transformers back to 4.x (Flax CLIP / T5). That is a step down from v26.5 (5.9.0). Transformer Engine is tagged rocm7.14.0 and matches the image ROCm, unlike v26.5 where the TE wheel was a rocm7.15 build on ROCm 7.14.0.

Primus source for v26.6

Use the release/v26.6 branch for both images:

git clone --recurse-submodules https://github.com/AMD-AGI/Primus.git
cd Primus
git checkout release/v26.6
git submodule update --init --recursive


Branch tip	`2aa05ead` (2026-08-25)
Megatron-LM	`d3528a21`
TorchTitan	`73a0e697`
Megatron Bridge	`9577b128`
MaxText	`b47d74bf`
Emerging-Optimizers	`93d9eb3a`
HummingbirdXT	`ed7b7bd0`

Prefer a release/v26.6 checkout over the Primus copy baked into the images.

rocm/primus:v26.6 was built from 2aa05ead (2026-08-25), which is the current release/v26.6 tip, so the in-image /workspace/Primus currently matches. Cloning the branch still keeps you current if later commits land on it.

rocm/jax-training:maxtext-v26.6 was built from main at 4d2f7a74 (2026-08-28), because its Dockerfile pins PRIMUS_BRANCH=main rather than a commit. Unlike v26.5, its bundled third_party/maxtext does match /workspace/maxtext (b47d74bf). Use release/v26.6 for MaxText training so you get the Primus recipes validated against this image rather than whatever main was on the build day.

Changes since v26.5

rocm/primus:

Component	v26.5	v26.6
ROCm nightly	7.15.0a20260720	7.15.0a20260727
PyTorch	2.12.0+rocm7.15.0a20260720	2.12.0+rocm7.15.0a20260727
Transformer Engine	2.15.0.dev0+rocm7.15.0a20260716.a07e607	2.17.0+rocm7.15.0a20260727.e028a6c
Flash Attention	2.8.3	2.8.1
Triton	3.7.1+git0263a6a6	3.8.0+git4cff872c
hipBLASLt	1.4.1-1aa46415	1.4.1-bbb68174
Primus-Turbo	0.3.2.dev48	0.4.1.dev26
APEX	1.14.0a0+rocm7.15.0a20260721	1.15.0a0+rocm10.1.0a20260822
transformers	4.55.0	5.5.0
FBGEMM	2026.7.22	2026.8.25
NumPy	2.5.1	2.5.2
Image size	54.7 GB	54.0 GB

rocm/jax-training:maxtext:

Component	v26.5	v26.6
JAX / jaxlib	0.10.0	0.11.0
jax-rocm7-pjrt / jax-rocm7-plugin	0.10.0+rocm7.14.0	0.11.0.post1
Transformer Engine	2.15.0.dev0+rocm7.15.0a20260707.72d01a0	2.17.0+rocm7.14.0.50a84ad
Flax	0.12.2	0.12.8
Orbax / Grain / tensorstore	0.11.39 / 0.2.16 / 0.1.82	0.12.4 / 0.2.18 / 0.1.85
MaxText	`a7c6c7e5`	`b47d74bf`
transformers	5.9.0	4.57.3
NumPy	2.0.2	2.5.2
Image size	45.7 GB	42.6 GB

JAX 0.11.0 still requires Shardy. Set shardy=True during the training run on v26.6. See the Shardy migration guide.

v26.5

`rocm/primus:v26.5`

Megatron-LM, TorchTitan, and Megatron Bridge backends.


Image ID	`3040bf42974d`
Built	2026-07-22
Size	54.7 GB
Manifest	`8e124a76fbe33cbcc26062f05da3ca5e6419b163`
Dockerfile	`Dockerfile.primus-v26.5`

Software component	Version
ROCm	7.15.0 (`rocm-sdk` 7.15.0a20260720)
Python	3.12.3
PyTorch	2.12.0+rocm7.15.0a20260720
Transformer Engine	2.15.0.dev0+rocm7.15.0a20260716.a07e607
Flash Attention	2.8.3
hipBLASLt	1.4.1-1aa46415
Triton	3.7.1+git0263a6a6.rocm7.15.0a20260720
RCCL	2.30.4
torchvision	0.27.0+rocm7.15.0a20260720
torchaudio	2.11.0+rocm7.15.0a20260721
APEX	1.14.0a0+rocm7.15.0a20260721
AITER	0.1.14.post1
Primus-Turbo	0.3.2.dev48
torchao	0.15.0+gite9c7bead9
FBGEMM	2026.7.22
mamba-ssm / causal-conv1d / grouped_gemm	2.3.1 / 1.5.0.post8 / 1.1.4
transformers / datasets	4.55.0 / 3.6.0
NumPy	2.5.1

`rocm/jax-training:maxtext-v26.5`

MaxText (JAX) backend.


Image ID	`b034a6769b58`
Built	2026-07-17
Size	45.7 GB
Manifest	`5ffe026e7bb969c42ce9f3c9f5c4a3c3164c8a8c`
Dockerfile	`Dockerfile.jax-v26.5`

Software component	Version
ROCm	7.14.0
Python	3.12.3
JAX / jaxlib	0.10.0
jax-rocm7-pjrt / jax-rocm7-plugin	0.10.0+rocm7.14.0
Transformer Engine	2.15.0.dev0+rocm7.15.0a20260707.72d01a0
hipBLASLt	1.4.1-cd957402
RCCL	2.30.4 (built from rocm-systems `9e5e4084`)
Flax	0.12.2
TensorFlow	2.21.0 (CPU-only, rebuilt from the ROCm fork)
Optax / Orbax / Grain / tensorstore	0.2.8 / 0.11.39 / 0.2.16 / 0.1.82
MaxText	`a7c6c7e5` (`release/v26.5`)
transformers / datasets	5.9.0 / 4.8.5
NumPy	2.0.2
amdsmi	7.0.2

Note: The Transformer Engine wheel is tagged rocm7.15.0a… even though the image ships ROCm 7.14.0. This is intentional and matches the Dockerfile.

Primus source for v26.5

Use the release/v26.5 branch for both images:

git clone --recurse-submodules https://github.com/AMD-AGI/Primus.git
cd Primus
git checkout release/v26.5
git submodule update --init --recursive


Branch tip	`ca7ddda3` (2026-07-30)
Megatron-LM	`d3528a21`
TorchTitan	`73a0e697`
Megatron Bridge	`9577b128`
MaxText	`a7c6c7e5`
Emerging-Optimizers	`93d9eb3a`
HummingbirdXT	`ed7b7bd0`

Do not use the Primus checkout baked into the images. Both images contain a /workspace/Primus, but neither tracks release/v26.5:

rocm/primus:v26.5 was built from b511d1b6 (2026-07-22). Three commits have landed on release/v26.5 since, which bumped the MaxText submodule from 80f431d0 to a7c6c7e5 and added support for the v26.5 MaxText API.

rocm/jax-training:maxtext-v26.5 was built from main at 8b5e8091 (2026-07-17), because its Dockerfile pins PRIMUS_BRANCH=main rather than a commit. Its bundled third_party/maxtext is 80f431d0, which does not match the /workspace/maxtext checkout (a7c6c7e5) the image was validated against.

Cloning release/v26.5 yourself avoids both problems and gives you the MaxText the image was built around.

Changes since v26.4

rocm/primus:

Component	v26.4	v26.5
ROCm	7.14.0	7.15.0
PyTorch	2.12.0+rocm7.14.0a20260608	2.12.0+rocm7.15.0a20260720
Transformer Engine	2.14.0.dev0+e6ede467	2.15.0.dev0+rocm7.15.0a20260716.a07e607
Triton	3.7.0+gitb4e20bbe	3.7.1+git0263a6a6
RCCL	2.29.7	2.30.4
hipBLASLt	1.4.1-be5adb9b	1.4.1-1aa46415
AITER	0.1.12.post2.dev214+gb5e03ed19	0.1.14.post1
Primus-Turbo	0.3.0+3c39ef2	0.3.2.dev48
APEX	1.11.0+rocm7.14.0a20260618	1.14.0a0+rocm7.15.0a20260721
NumPy	2.4.6	2.5.1
Image size	75.9 GB	54.7 GB

rocm/jax-training:maxtext:

Component	v26.4	v26.5
JAX / jaxlib	0.9.1	0.10.0
jax-rocm7-pjrt / jax-rocm7-plugin	0.9.1+rocm7.14.0a20260526	0.10.0+rocm7.14.0
Transformer Engine	2.12.0.dev0+635d7c08	2.15.0.dev0+rocm7.15.0a20260707.72d01a0
hipBLASLt	1.4.0-807283e5	1.4.1-cd957402
RCCL	2.28.9	2.30.4
TensorFlow	2.20.0	2.21.0
MaxText	`80f431d0`	`a7c6c7e5`
transformers	5.8.1	5.9.0
Image size	63.6 GB	45.7 GB

JAX 0.10.0 requires Shardy. Set shardy=True during the training run on v26.5. See the Shardy migration guide.

v26.4

`rocm/primus:v26.4`


Image ID	`8c8ecc6fe14b`
Built	2026-06-18
Size	75.9 GB
Manifest	`a2d4c7239d874501f864080baf2a8131236b0221`
Also published as	`rocm/primus:pytorch-2.12.0-rocm7.14.0a20260608_te-2.14.0.dev0-e6ede467_v26.4` (same image)
Dockerfile	`Dockerfile.primus-v26.4`

Software component	Version
ROCm	7.14.0 (`rocm-sdk` 7.14.0a20260608)
Python	3.12.3
PyTorch	2.12.0+rocm7.14.0a20260608
Transformer Engine	2.14.0.dev0+e6ede467
Flash Attention	2.8.3
hipBLASLt	1.4.1-be5adb9b
Triton	3.7.0+gitb4e20bbe.rocm7.14.0a20260608
RCCL	2.29.7
torchvision	0.27.0+rocm7.14.0a20260608
torchaudio	2.11.0+rocm7.14.0a20260617
APEX	1.11.0+rocm7.14.0a20260618
AITER	0.1.12.post2.dev214+gb5e03ed19
Primus-Turbo	0.3.0+3c39ef2
torchao	0.15.0+gite9c7bead9
FBGEMM	2026.6.18
mamba-ssm / causal-conv1d / grouped_gemm	2.3.1 / 1.5.0.post8 / 1.1.4
transformers / datasets	4.55.0 / 3.6.0
NumPy	2.4.6

`rocm/jax-training:maxtext-v26.4`


Image ID	`9b263b38c159`
Built	2026-05-27
Size	63.6 GB
Manifest	`b0b705ce0533ead5490070daa402ca625f1489b5`
Also published as	`rocm/jax-training:maxtext-v26.4-jax0.9.1-te2.12.0` (same image)
Dockerfile	`Dockerfile.jax-v26.4`

Software component	Version
ROCm	7.14.0
Python	3.12.3
JAX / jaxlib	0.9.1
jax-rocm7-pjrt / jax-rocm7-plugin	0.9.1+rocm7.14.0a20260526
Transformer Engine	2.12.0.dev0+635d7c08
hipBLASLt	1.4.0-807283e5
RCCL	2.28.9
Flax	0.12.2
TensorFlow	2.20.0 (CPU-only)
Optax / Orbax / Grain / tensorstore	0.2.8 / 0.11.39 / 0.2.16 / 0.1.82
MaxText	`80f431d0`
transformers / datasets	5.8.1 / 4.8.5
NumPy	2.0.2
amdsmi	7.0.2

Primus source for v26.4

rocm/primus:v26.4 bakes Primus at 236cfa9d, which is the commit the published v26.4 benchmark recipes pin:

git clone --recurse-submodules https://github.com/AMD-AGI/Primus.git
cd Primus
git checkout 236cfa9d
git submodule update --init --recursive

Earlier releases

Headline versions only. Read the in-image manifest for the full stack of any image below — except v26.2 and v26.1, which predate the manifest (use pip list there).

The three MaxText v26.3.x images ship an identical software stack; they differ only in MaxText and Primus content.

Image	Python	ROCm	Framework	Transformer Engine	RCCL
`rocm/primus:v26.3`	3.12.3	7.2.1	PyTorch 2.10.0+git94c6e04	2.12.0.dev0+40434cf6	2.27.7
`rocm/primus:v26.2`	3.12.3	7.2.0	PyTorch 2.10.0a0+git449b176	2.8.0.dev0+51f74fa7	2.27.7
`rocm/primus:v26.1`	3.10.12	7.1.0	PyTorch 2.10.0.dev20251112+rocm7.1	2.6.0.dev0+f141f34b	2.27.7
`rocm/jax-training:maxtext-v26.3.2`	3.12.3	7.2.1	JAX 0.8.2	2.8.0.dev0+9b312832	2.27.7
`rocm/jax-training:maxtext-v26.3.1`	3.12.3	7.2.1	JAX 0.8.2	2.8.0.dev0+9b312832	2.27.7
`rocm/jax-training:maxtext-v26.3`	3.12.3	7.2.1	JAX 0.8.2	2.8.0.dev0+9b312832	2.27.7
`rocm/jax-training:maxtext-v26.2`	3.12.3	7.1.1	JAX 0.8.2	2.8.0.dev0+aec00a7f	2.27.7

Verifying the stack in an image

Images from v26.3 onward ship a manifest at /workspace/.manifest/ recording exactly what was installed at build time:

File	Contents
`requirements.txt`	full `pip list`
`dpkg-list.txt`	full `dpkg -l`
`env.txt`	every environment variable baked into the image
`training_docker_version`	the build's commit tag
`Dockerfile`	the Dockerfile the image was built from (`docker-build-recipe.txt` in `rocm/jax-training:maxtext-v26.6`)

docker run --rm --entrypoint bash rocm/primus:v26.6 -c 'cat /workspace/.manifest/requirements.txt'

Native library versions are not pip packages; read them from the ROCm headers:

docker run --rm --entrypoint bash rocm/primus:v26.6 -c '
  grep -E "HIPBLASLT_VERSION_(MAJOR|MINOR|PATCH|TWEAK)" $(find $ROCM_PATH /opt/rocm -name hipblaslt-version.h 2>/dev/null | head -1)
  grep -E "define NCCL_(MAJOR|MINOR|PATCH)"              $(find $ROCM_PATH /opt/rocm -name rccl.h            2>/dev/null | head -1)'

Images older than v26.3 predate the manifest; query them with pip list directly.

Navigation Menu

Repository navigation

release-notes.md

release-notes.md

Release notes

v26.6 (current)

`rocm/primus:v26.6`

`rocm/jax-training:maxtext-v26.6`

Primus source for v26.6

Changes since v26.5

v26.5

`rocm/primus:v26.5`

`rocm/jax-training:maxtext-v26.5`

Primus source for v26.5

Changes since v26.4

v26.4

`rocm/primus:v26.4`

`rocm/jax-training:maxtext-v26.4`

Primus source for v26.4

Earlier releases

Verifying the stack in an image

Related documentation

FilesExpand file tree

release-notes.md

Latest commit

History

release-notes.md

File metadata and controls

Release notes

v26.6 (current)

rocm/primus:v26.6

rocm/jax-training:maxtext-v26.6

Primus source for v26.6

Changes since v26.5

v26.5

rocm/primus:v26.5

rocm/jax-training:maxtext-v26.5

Primus source for v26.5

Changes since v26.4

v26.4

rocm/primus:v26.4

rocm/jax-training:maxtext-v26.4

Primus source for v26.4

Earlier releases

Verifying the stack in an image

Related documentation

`rocm/primus:v26.6`

`rocm/jax-training:maxtext-v26.6`

`rocm/primus:v26.5`

`rocm/jax-training:maxtext-v26.5`

`rocm/primus:v26.4`

`rocm/jax-training:maxtext-v26.4`