Release notes

September 17, 2026 · View on GitHub

AMD publishes two families of Primus training Docker image:

ImageBackendsDocumented in
rocm/primus:<version>Megatron-LM, TorchTitan, Megatron BridgeMegatron-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.

Highlights for v26.7

ROCm 10.0.0. Both image families move off the ROCm 7.x line. This is the defining change of the release and it reaches everything: PyTorch, Transformer Engine, Triton, torchvision/torchaudio and APEX are all rebuilt against it, and the JAX plugin pair is renamed jax-rocm10-pjrt / jax-rocm10-plugin. If you install bare metal, note that the wheel indexes moved too — see Bare-metal installation.

Beyond that, the backend work in this release is weighted towards the PyTorch family; MaxText picks up the ROCm 10 stack plus config tuning rather than new capability.

Both families

  • gfx1250 multi-GPU enablement. Primus recognises gfx1250 by ISA rather than by MI number and ships an arch env file for it, so primus-cli applies real settings instead of falling through to Detected GPU model: unknown (#1043).
  • primus-cli honours an explicit flag whose value happens to equal the parser default (#1002).
  • *_SOCKET_IFNAME no longer falls back to an IP address (#1063).
  • The primus wheel builds again. The mlperf-logging / mlperf-common VCS pins moved to an optional mlperf extra, because Hatchling 1.32 rejects PEP 508 direct references in project.dependencies (#1061).

rocm/primus — Megatron-LM and TorchTitan

New features

  • DeepSeek-V4 on gfx942 at long context. V4 upstream targeted gfx950 and in practice only ran at seq_length=4096; it now trains on MI308X/CDNA3 with context parallelism at 128k, full model (#1047).
  • Packed-sequence (THD) SFT for DeepSeek-V4 — many samples per window delimited by cu_seqlens, with attention isolated so no query sees another sample; full model at 4k and 128k across three nodes (#1062).
  • DLRM-v4 (TorchRec/HSTU) projection workload with first-principles MI350X calibration (#1059).
  • MLPerf Llama 3.1 8B launcher and refreshed MI355X benchmark configs (#1012).

Performance

  • Fused SwiGLU + fc2 (opt-in, via FLA swiglu_linear) removes one FFN-wide saved tensor per MLP layer by recomputing the activation in the backward pass (#1051).
  • Fused cross-entropy no longer copies the full logits tensor (#1049).
  • DeepSeek-V4 indexer distillation loss is fused and enabled (#992).
  • Tuned configs: GDN/KDA 1B on MI355X for higher occupancy (#1080), and MI355X Llama 3.1 8B / GPT-OSS 20B defaults (#1045).

Bug fixes

  • Weight gradients were dropped after the first microbatch on Turbo's non-fused _bridge_weight_grad path, which gated accumulation on grad_added_to_main_grad. Any run using gradient accumulation on that path was training on partial gradients (#1046).
  • Out-of-bounds top-k index in the V4 DSA forward kernel is now bounded (#1044).
  • fused_softcap stays finite on saturation-range logits (#1057).
  • KDA fused in_proj is padded past the hipBLASLt bf16 dead zone (#1050).
  • Pipeline warmup is clamped for short batches (#1032).
  • TorchTitan Turbo grouped-GEMM config renamed (#1041). Update any local config that sets the old key.

rocm/jax-training:maxtext — MaxText

No new MaxText capability this release; the change is the ROCm 10.0.0 stack itself, plus config tuning:

  • MI300X batch-size tuning across the MaxText recipes (#1020).
  • gemma4_26B-fp8 retuned and the pure_nnx_decoder fp8 workaround dropped, now that it is no longer needed (#1019).

v26.7 (current)

rocm/primus:v26.7

Megatron-LM, TorchTitan, and Megatron Bridge backends.

Image ID68b7eb7d4db9
Built2026-09-09
Size53.8 GB
Manifestdb0de753f37d6a6b782bc836b2be54d8ae362258
DockerfileDockerfile.primus-v26.7
Software componentVersion
ROCm10.0.0
Python3.12.3
PyTorch2.12.0+rocm10.0.0
Transformer Engine2.17.0+rocm10.0.0
Flash Attention2.8.1
hipBLASLt1.4.1-8d1ae90e
Triton3.8.0+git4cff872c.rocm10.0.0
RCCL2.30.4
torchvision0.27.0+rocm10.0.0
torchaudio2.11.0+rocm10.0.0
APEX1.13.0+rocm10.0.0
AITER0.1.14.post1
Primus-Turbo0.4.1.dev33
torchao0.15.0+gite9c7bead9
FBGEMM2026.9.9
mamba-ssm / causal-conv1d / grouped_gemm2.3.1 / 1.5.0.post8 / 1.1.4
transformers / datasets5.10.0 / 3.6.0
NumPy2.5.3

rocm/jax-training:maxtext-v26.7

MaxText (JAX) backend.

Image IDb5117e775594
Built2026-09-09
Size43.7 GB
Manifest8da24470ccb3b16c60ddb6f8dfc1a1526bc2a408
DockerfileDockerfile.jax-v26.7
Software componentVersion
ROCm10.0.0
Python3.12.3
JAX / jaxlib0.11.0
jax-rocm10-pjrt / jax-rocm10-plugin0.11.0+rocm10.0.0
Transformer Engine2.17.0+rocm10.0.0
hipBLASLt1.4.1-8d1ae90e
RCCL2.30.4
Flax0.12.8
TensorFlow2.21.0
Optax / Orbax / Grain / tensorstore0.2.8 / 0.12.4 / 0.2.18 / 0.1.85
MaxTextb3c53763
transformers / datasets4.57.3 / 4.8.5
NumPy2.5.3

ROCm 10.0.0 changes where the wheels come from. Both families now install ROCm core from stable.repo.amd.com/rocm/core/whl-next, with PyTorch, Transformer Engine and the JAX plugin on sibling indexes. The rocm.nightlies.amd.com/whl-multi-arch index used through v26.6 does not carry these wheels. Transformer Engine also ships as three distributions (transformer_engine, transformer_engine_rocm10, and the torch or jax flavour) where v26.6 had two.

RCCL is no longer rebuilt. v26.6 overrode the SDK's RCCL with a rocm-systems build; v26.7 uses RCCL 2.30.4 as shipped in the ROCm 10.0.0 SDK. The version is unchanged, the provenance is not.

Two rows in the delta table need reading carefully. APEX appears to go backwards, from 1.15.0a0+rocm10.1.0a20260822 to 1.13.0+rocm10.0.0: v26.6 pulled a rocm10.1 nightly onto a ROCm 7.15 base, whereas v26.7 uses the APEX built against its own ROCm 10.0.0. And the JAX plugin row is labelled with the v26.7 spelling on both sides — the v26.6 value 0.11.0.post1 shipped as jax-rocm7-pjrt / jax-rocm7-plugin, before the rename.

Note: transformers is 4.57.3 in the MaxText image because the published tag includes the MaxDiffusion stage, which pins transformers back to 4.x for Flax CLIP / T5. The build-time manifest records 5.14.1, captured before that stage runs; 4.57.3 is what you actually get. The PyTorch image is on 5.10.0.

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

Primus source for v26.7

Use the release/v26.7 branch for both images:

git clone --recurse-submodules https://github.com/AMD-AGI/Primus.git
cd Primus
git checkout release/v26.7
git submodule update --init --recursive
Branch tip2631e68d (2026-09-02), which is also the v26.7.0 tag
Megatron-LMd3528a21
TorchTitan73a0e697
Megatron Bridge9577b128
MaxTextb3c53763
Emerging-Optimizers93d9eb3a
HummingbirdXTed7b7bd0

Prefer a release/v26.7 checkout over the Primus copy baked into the images.

  • rocm/primus:v26.7 was built from 2631e68d, which is the current release/v26.7 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-v26.7 was built from main at e7968675 (2026-09-09), because its Dockerfile pins PRIMUS_BRANCH=main rather than a commit. Use release/v26.7 anyway, so you get the Primus recipes the release was validated against rather than whatever main held on the build day. Its bundled third_party/maxtext matches /workspace/maxtext (b3c53763).

Changes since v26.6

rocm/primus:

Componentv26.6v26.7
ROCm7.15.0a2026072710.0.0
PyTorch2.12.0+rocm7.15.0a202607272.12.0+rocm10.0.0
Transformer Engine2.17.0+rocm7.15.0a20260727.e028a6c2.17.0+rocm10.0.0
hipBLASLt1.4.1-bbb681741.4.1-8d1ae90e
Triton3.8.0+git4cff872c.rocm7.15.0a202607273.8.0+git4cff872c.rocm10.0.0
torchvision0.27.0+rocm7.15.0a202607270.27.0+rocm10.0.0
torchaudio2.11.0+rocm7.15.0a202607282.11.0+rocm10.0.0
APEX1.15.0a0+rocm10.1.0a202608221.13.0+rocm10.0.0
Primus-Turbo0.4.1.dev260.4.1.dev33
FBGEMM2026.8.252026.9.9
transformers / datasets5.5.0 / 3.6.05.10.0 / 3.6.0
NumPy2.5.22.5.3
Image size54.0 GB53.8 GB

rocm/jax-training:maxtext:

Componentv26.6v26.7
ROCm7.14.010.0.0
jax-rocm10-pjrt / jax-rocm10-plugin0.11.0.post10.11.0+rocm10.0.0
Transformer Engine2.17.0+rocm7.14.0.50a84ad2.17.0+rocm10.0.0
hipBLASLt1.4.1-cd9574021.4.1-8d1ae90e
MaxTextb47d74bfb3c53763
NumPy2.5.22.5.3
Image size42.6 GB43.7 GB

v26.6

rocm/primus:v26.6

Megatron-LM, TorchTitan, and Megatron Bridge backends.

Image ID4fcb3f210dc6
Built2026-08-25
Size54.0 GB
Manifestf756f2279d8ab57b6549bcbd50d249755b69a407
DockerfileDockerfile.primus-v26.6
Software componentVersion
ROCm7.15.0 (rocm-sdk 7.15.0a20260727)
Python3.12.3
PyTorch2.12.0+rocm7.15.0a20260727
Transformer Engine2.17.0+rocm7.15.0a20260727.e028a6c
Flash Attention2.8.1
hipBLASLt1.4.1-bbb68174
Triton3.8.0+git4cff872c.rocm7.15.0a20260727
RCCL2.30.4
torchvision0.27.0+rocm7.15.0a20260727
torchaudio2.11.0+rocm7.15.0a20260728
APEX1.15.0a0+rocm10.1.0a20260822
AITER0.1.14.post1
Primus-Turbo0.4.1.dev26
torchao0.15.0+gite9c7bead9
FBGEMM2026.8.25
mamba-ssm / causal-conv1d / grouped_gemm2.3.1 / 1.5.0.post8 / 1.1.4
transformers / datasets5.5.0 / 3.6.0
NumPy2.5.2

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 IDa71d8dbb045e
Built2026-08-28
Size42.6 GB
Manifest31a3a21d0a37cbd0b5de0342535f65d557ff77ba
DockerfileDockerfile.jax-v26.6
Software componentVersion
ROCm7.14.0
Python3.12.3
JAX / jaxlib0.11.0
jax-rocm7-pjrt / jax-rocm7-plugin0.11.0.post1
Transformer Engine2.17.0+rocm7.14.0.50a84ad
hipBLASLt1.4.1-cd957402
RCCL2.30.4 (built from rocm-systems 9e5e4084)
Flax0.12.8
TensorFlow2.21.0 (CPU-only, rebuilt from the ROCm fork)
Optax / Orbax / Grain / tensorstore0.2.8 / 0.12.4 / 0.2.18 / 0.1.85
MaxTextb47d74bf (release/v26.6)
transformers / datasets4.57.3 / 4.8.5
NumPy2.5.2

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 tip2aa05ead (2026-08-25)
Megatron-LMd3528a21
TorchTitan73a0e697
Megatron Bridge9577b128
MaxTextb47d74bf
Emerging-Optimizers93d9eb3a
HummingbirdXTed7b7bd0

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:

Componentv26.5v26.6
ROCm nightly7.15.0a202607207.15.0a20260727
PyTorch2.12.0+rocm7.15.0a202607202.12.0+rocm7.15.0a20260727
Transformer Engine2.15.0.dev0+rocm7.15.0a20260716.a07e6072.17.0+rocm7.15.0a20260727.e028a6c
Flash Attention2.8.32.8.1
Triton3.7.1+git0263a6a63.8.0+git4cff872c
hipBLASLt1.4.1-1aa464151.4.1-bbb68174
Primus-Turbo0.3.2.dev480.4.1.dev26
APEX1.14.0a0+rocm7.15.0a202607211.15.0a0+rocm10.1.0a20260822
transformers4.55.05.5.0
FBGEMM2026.7.222026.8.25
NumPy2.5.12.5.2
Image size54.7 GB54.0 GB

rocm/jax-training:maxtext:

Componentv26.5v26.6
JAX / jaxlib0.10.00.11.0
jax-rocm7-pjrt / jax-rocm7-plugin0.10.0+rocm7.14.00.11.0.post1
Transformer Engine2.15.0.dev0+rocm7.15.0a20260707.72d01a02.17.0+rocm7.14.0.50a84ad
Flax0.12.20.12.8
Orbax / Grain / tensorstore0.11.39 / 0.2.16 / 0.1.820.12.4 / 0.2.18 / 0.1.85
MaxTexta7c6c7e5b47d74bf
transformers5.9.04.57.3
NumPy2.0.22.5.2
Image size45.7 GB42.6 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 ID3040bf42974d
Built2026-07-22
Size54.7 GB
Manifest8e124a76fbe33cbcc26062f05da3ca5e6419b163
DockerfileDockerfile.primus-v26.5
Software componentVersion
ROCm7.15.0 (rocm-sdk 7.15.0a20260720)
Python3.12.3
PyTorch2.12.0+rocm7.15.0a20260720
Transformer Engine2.15.0.dev0+rocm7.15.0a20260716.a07e607
Flash Attention2.8.3
hipBLASLt1.4.1-1aa46415
Triton3.7.1+git0263a6a6.rocm7.15.0a20260720
RCCL2.30.4
torchvision0.27.0+rocm7.15.0a20260720
torchaudio2.11.0+rocm7.15.0a20260721
APEX1.14.0a0+rocm7.15.0a20260721
AITER0.1.14.post1
Primus-Turbo0.3.2.dev48
torchao0.15.0+gite9c7bead9
FBGEMM2026.7.22
mamba-ssm / causal-conv1d / grouped_gemm2.3.1 / 1.5.0.post8 / 1.1.4
transformers / datasets4.55.0 / 3.6.0
NumPy2.5.1

rocm/jax-training:maxtext-v26.5

MaxText (JAX) backend.

Image IDb034a6769b58
Built2026-07-17
Size45.7 GB
Manifest5ffe026e7bb969c42ce9f3c9f5c4a3c3164c8a8c
DockerfileDockerfile.jax-v26.5
Software componentVersion
ROCm7.14.0
Python3.12.3
JAX / jaxlib0.10.0
jax-rocm7-pjrt / jax-rocm7-plugin0.10.0+rocm7.14.0
Transformer Engine2.15.0.dev0+rocm7.15.0a20260707.72d01a0
hipBLASLt1.4.1-cd957402
RCCL2.30.4 (built from rocm-systems 9e5e4084)
Flax0.12.2
TensorFlow2.21.0 (CPU-only, rebuilt from the ROCm fork)
Optax / Orbax / Grain / tensorstore0.2.8 / 0.11.39 / 0.2.16 / 0.1.82
MaxTexta7c6c7e5 (release/v26.5)
transformers / datasets5.9.0 / 4.8.5
NumPy2.0.2
amdsmi7.0.2

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 tipca7ddda3 (2026-07-30)
Megatron-LMd3528a21
TorchTitan73a0e697
Megatron Bridge9577b128
MaxTexta7c6c7e5
Emerging-Optimizers93d9eb3a
HummingbirdXTed7b7bd0

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:

Componentv26.4v26.5
ROCm7.14.07.15.0
PyTorch2.12.0+rocm7.14.0a202606082.12.0+rocm7.15.0a20260720
Transformer Engine2.14.0.dev0+e6ede4672.15.0.dev0+rocm7.15.0a20260716.a07e607
Triton3.7.0+gitb4e20bbe3.7.1+git0263a6a6
RCCL2.29.72.30.4
hipBLASLt1.4.1-be5adb9b1.4.1-1aa46415
AITER0.1.12.post2.dev214+gb5e03ed190.1.14.post1
Primus-Turbo0.3.0+3c39ef20.3.2.dev48
APEX1.11.0+rocm7.14.0a202606181.14.0a0+rocm7.15.0a20260721
NumPy2.4.62.5.1
Image size75.9 GB54.7 GB

rocm/jax-training:maxtext:

Componentv26.4v26.5
JAX / jaxlib0.9.10.10.0
jax-rocm7-pjrt / jax-rocm7-plugin0.9.1+rocm7.14.0a202605260.10.0+rocm7.14.0
Transformer Engine2.12.0.dev0+635d7c082.15.0.dev0+rocm7.15.0a20260707.72d01a0
hipBLASLt1.4.0-807283e51.4.1-cd957402
RCCL2.28.92.30.4
TensorFlow2.20.02.21.0
MaxText80f431d0a7c6c7e5
transformers5.8.15.9.0
Image size63.6 GB45.7 GB

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

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.

ImagePythonROCmFrameworkTransformer EngineRCCL
rocm/primus:v26.43.12.37.14.0PyTorch 2.12.0+rocm7.14.0a202606082.14.0.dev0+e6ede4672.29.7
rocm/primus:v26.33.12.37.2.1PyTorch 2.10.0+git94c6e042.12.0.dev0+40434cf62.27.7
rocm/primus:v26.23.12.37.2.0PyTorch 2.10.0a0+git449b1762.8.0.dev0+51f74fa72.27.7
rocm/primus:v26.13.10.127.1.0PyTorch 2.10.0.dev20251112+rocm7.12.6.0.dev0+f141f34b2.27.7
rocm/jax-training:maxtext-v26.43.12.37.14.0JAX 0.9.12.12.0.dev0+635d7c082.28.9
rocm/jax-training:maxtext-v26.3.23.12.37.2.1JAX 0.8.22.8.0.dev0+9b3128322.27.7
rocm/jax-training:maxtext-v26.3.13.12.37.2.1JAX 0.8.22.8.0.dev0+9b3128322.27.7
rocm/jax-training:maxtext-v26.33.12.37.2.1JAX 0.8.22.8.0.dev0+9b3128322.27.7
rocm/jax-training:maxtext-v26.23.12.37.1.1JAX 0.8.22.8.0.dev0+aec00a7f2.27.7

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:

FileContents
requirements.txtfull pip list
dpkg-list.txtfull dpkg -l
env.txtevery environment variable baked into the image
training_docker_versionthe build's commit tag
Dockerfilethe 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.