Added Dockerfile for CI images & Upgrate CI to ROCm 7.2

[VeeraRajasekhar]

Description

Added the dockerfile, which can be used to create the ci-artifactory images.

Fixes # (issue)

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

Please list the changes introduced in this PR:

• Added a new file docker/Dockerfile

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[wenchenvincent]

Please address the comments.

[ipanfilo]

Why conversations are marked as resolved w/o any actual action?

[VeeraRajasekhar]

Why conversations are marked as resolved w/o any actual action?

Some of them, I have resolved, some I have currently resolved in my local, just to keep track I will mark them resolved.

[wenchenvincent]

@VeeraRajasekhar Is this PR still needed?

[wenchenvincent]

@VeeraRajasekhar Could you remind me of what we had decided on this PR? It seemed that it is no longer relevant and we should close it.

[VeeraRajasekhar]

Hi @ipanfilo, @wangye805

I have updated this PR with latest 7.2 docker file and moved to .github/scripts.

Let me know if I need to add an action to automate docker build and upload to our artifactory?

Thanks.

[VeeraRajasekhar]

I had to force push to include new FA 2.8.3 support commit and my changes for 7.2 support to run the CI.

Thanks.

[VeeraRajasekhar]

@Micky774, please review the following,

Analysis on testing on Jax & xla 0.8.2

(Not Supported) jax.nn.scaled_matmul (MXFP8) on ROCm crashes with a segmentation fault when the contracting dimension (K) is less than 64.

import functools
import jax
import jax.numpy as jnp
from jax import nn

key = jax.random.PRNGKey(0)
key_a, key_b = jax.random.split(key)
B, M, N, K = 1, 128, 128, 32


lhs = jax.random.normal(key_a, (B, M, K), dtype=jnp.float32)
rhs = jax.random.normal(key_b, (B, N, K), dtype=jnp.float32)

# 1. high-precision matmul
ref = jnp.einsum("bmk,bnk->bmn", lhs, rhs)

# 2. mxfp8 matmul
configs = [nn.get_scaled_dot_general_config("mxfp8")] * 3
scaled_dot = functools.partial(
    nn.scaled_dot_general,
    configs=configs,
    preferred_element_type=jnp.float32,
)

out = scaled_dot(lhs, rhs, (((2,), (2,)), ((0,), (0,))))

# compare results
print("high-precision ref: ")
print(ref)

print("mxfp8 out: ")
print(out)

max_abs = jnp.max(jnp.abs(out - ref))
max_rel = max_abs / jnp.max(jnp.abs(ref))

print("max abs error:", max_abs)
print("max rel error:", max_rel)

[rr/error] Predicate mismatch for RR_GEMM_TN_FP8_FP8_Float_Float_Float_SA_B_SB_B_WGT_32x32x64_UR_2: K must be a multiple of workgroupTile.k=64 * unrollK=1
[rr/error] Predicate mismatch for RR_GEMM_TN_FP8_FP8_Float_Float_Float_SA_B_SB_B_WGT_16x16x128_UR_2: K must be a multiple of workgroupTile.k=128 * unrollK=1
[rr/error] Predicate mismatch for RR_GEMM_TN_FP8_FP8_Float_Float_Float_SA_B_SB_B_WGT_32x16x128_UR_2: K must be a multiple of workgroupTile.k=128 * unrollK=1
[rr/error] Predicate mismatch for RR_GEMM_TN_FP8_FP8_Float_Float_Float_SA_B_SB_B_WGT_16x32x128_UR_2: K must be a multiple of workgroupTile.k=128 * unrollK=1
[rr/error] Predicate mismatch for RR_GEMM_TN_FP8_FP8_Float_Float_Float_SA_B_SB_B_WGT_32x32x128_UR_2: K must be a multiple of workgroupTile.k=128 * unrollK=1
[rr/error] Predicate mismatch for RR_GEMM_TN_FP8_FP8_Float_Float_Float_SA_B_SB_B_WGT_64x16x128_UR_2: K must be a multiple of workgroupTile.k=128 * unrollK=1
[rr/error] Predicate mismatch for RR_GEMM_TN_FP8_FP8_Float_Float_Float_SA_B_SB_B_WGT_16x64x128_UR_2: K must be a multiple of workgroupTile.k=128 * unrollK=1
[rr/error] Predicate mismatch for RR_GEMM_TN_FP8_FP8_Float_Float_Float_SA_B_SB_B_WGT_64x32x128_UR_2: K must be a multiple of workgroupTile.k=128 * unrollK=1
[rr/error] Predicate mismatch for RR_GEMM_TN_FP8_FP8_Float_Float_Float_SA_B_SB_B_WGT_32x64x128_UR_2: K must be a multiple of workgroupTile.k=128 * unrollK=1
[rr/error] Predicate mismatch for RR_GEMM_TN_FP8_FP8_Float_Float_Float_SA_B_SB_B_WGT_16x16x256_UR_2: K must be a multiple of workgroupTile.k=256 * unrollK=1
[rr/error] Predicate mismatch for RR_GEMM_TN_FP8_FP8_Float_Float_Float_SA_B_SB_B_WGT_64x64x128_UR_2: K must be a multiple of workgroupTile.k=128 * unrollK=1
[rr/error] Predicate mismatch for RR_GEMM_TN_FP8_FP8_Float_Float_Float_SA_B_SB_B_WGT_128x32x128_UR_2: K must be a multiple of workgroupTile.k=128 * unrollK=1
[rr/error] Predicate mismatch for RR_GEMM_TN_FP8_FP8_Float_Float_Float_SA_B_SB_B_WGT_32x128x128_UR_2: K must be a multiple of workgroupTile.k=128 * unrollK=1
[rr/error] Predicate mismatch for RR_GEMM_TN_FP8_FP8_Float_Float_Float_SA_B_SB_B_WGT_128x64x128_UR_2: K must be a multiple of workgroupTile.k=128 * unrollK=1
[rr/error] Predicate mismatch for RR_GEMM_TN_FP8_FP8_Float_Float_Float_SA_B_SB_B_WGT_64x128x128_UR_2: K must be a multiple of workgroupTile.k=128 * unrollK=1
[rr/error] Predicate mismatch for RR_GEMM_TN_FP8_FP8_Float_Float_Float_SA_B_SB_B_WGT_16x64x256_UR_2: K must be a multiple of workgroupTile.k=256 * unrollK=1
[rr/error] Predicate mismatch for RR_GEMM_TN_FP8_FP8_Float_Float_Float_SA_B_SB_B_WGT_192x32x128_UR_2: M must be a multiple of workgroupTile.m=192
[rr/error] Predicate mismatch for RR_GEMM_TN_FP8_FP8_Float_Float_Float_SA_B_SB_B_WGT_192x32x128_UR_2: K must be a multiple of workgroupTile.k=128 * unrollK=1
[rr/error] Predicate mismatch for RR_GEMM_TN_FP8_FP8_Float_Float_Float_SA_B_SB_B_WGT_128x128x128_UR_2: K must be a multiple of workgroupTile.k=128 * unrollK=1
[rr/error] Predicate mismatch for RR_GEMM_TN_FP8_FP8_Float_Float_Float_SA_B_SB_B_WGT_32x192x128_UR_2: N must be a multiple of workgroupTile.n=192
[rr/error] Predicate mismatch for RR_GEMM_TN_FP8_FP8_Float_Float_Float_SA_B_SB_B_WGT_32x192x128_UR_2: K must be a multiple of workgroupTile.k=128 * unrollK=1
[rr/error] Predicate mismatch for RR_GEMM_TN_FP8_FP8_Float_Float_Float_SA_B_SB_B_WGT_192x64x128_UR_2: M must be a multiple of workgroupTile.m=192
[rr/error] Predicate mismatch for RR_GEMM_TN_FP8_FP8_Float_Float_Float_SA_B_SB_B_WGT_192x64x128_UR_2: K must be a multiple of workgroupTile.k=128 * unrollK=1
[rr/error] Predicate mismatch for RR_GEMM_TN_FP8_FP8_Float_Float_Float_SA_B_SB_B_WGT_64x192x128_UR_2: N must be a multiple of workgroupTile.n=192
[rr/error] Predicate mismatch for RR_GEMM_TN_FP8_FP8_Float_Float_Float_SA_B_SB_B_WGT_64x192x128_UR_2: K must be a multiple of workgroupTile.k=128 * unrollK=1
Segmentation fault (core dumped)

[VeeraRajasekhar]

https://github.com/ROCm/TransformerEngine/actions/runs/22024090773/job/63637830333

Level=3 testing had no issues.

[Micky774]

@Micky774, please review the following,

Analysis on testing on Jax & xla 0.8.2

(Not Supported) jax.nn.scaled_matmul (MXFP8) on ROCm crashes with a segmentation fault when the contracting dimension (K) is less than 64.
...

This is a failure on certain configs for hipblaslt, which we already have tickets open for. We don't support these configs in TE anyways, so it's a known issue and not a blocker. Shouldn't be a problem for this PR.

[Micky774]

@VeeraRajasekhar you'll need to merge w/ dev to fix CI

[VeeraRajasekhar]

https://github.com/ROCm/TransformerEngine/actions/runs/22285545864/job/64463404147