Add mi300 runner for toy_llama tests

[stbaione]

Currently, we only run on a w7900 when we run the toy_llama tests. I recently bisected an issue due to sharding accuracy being broken for llama in shortfin, on mi300 (#19872), and when I run iree-test-suites locally, I get failures:

================================================= short test summary info ==================================================
PASSED sharktank_models/clip/test_clip.py::test_results_close[bf16-hip]
PASSED sharktank_models/clip/test_clip.py::test_results_close[f32-hip]
PASSED sharktank_models/llama3.1/test_llama.py::test_prefill[hip]
PASSED sharktank_models/llama3.1/test_llama.py::test_decode[hip]
PASSED sharktank_models/llama3.1/test_llama.py::test_prefill_decode[hip]
PASSED sharktank_models/llama3.1/test_llama.py::test_prefill[hip_tp2]
PASSED sharktank_models/llama3.1/test_llama.py::test_prefill_decode[hip_tp2]
FAILED sharktank_models/llama3.1/test_llama.py::test_decode[hip_tp2] - AssertionError: cross entropy outside of tolerance
================================== 1 failed, 7 passed, 8 deselected, 2 warnings in 17.84s ==================================

I suspect that this would have been able to catch the bug had it been enabled. If it didn't, then we can determine how to extend the tests to cover the case we saw.

Reference issues for sharded failure: #19948, nod-ai/shark-ai#934

ci-exactly: build_packages,test_sharktank

[ScottTodd]

Looks like these runners need GPU drivers, either preinstalled or via Docker?

https://github.com/iree-org/iree/actions/runs/13270996475/job/37050767466?pr=19961#step:8:330

ERROR iree-test-suites/sharktank_models/llama3.1/test_llama.py::test_prefill[hip] - RuntimeError: Error creating driver: iree/runtime/src/iree/hal/drivers/hip/dynamic_symbols.c:160: UNAVAILABLE; HIP runtime library 'amdhip64.dll'/'libamdhip64.so' not available: please ensure installed and in dynamic library search path: 
  Tried: libamdhip64.so
    iree/runtime/src/iree/base/internal/dynamic_library_posix.c:165: NOT_FOUND; failed to load dynamic library (possibly not found on any search path): libamdhip64.so: cannot open shared object file: No such file or directory

See what this other workflow does:

iree/.github/workflows/pkgci_test_amd_mi300.yml

Lines 20 to 25 in e5a5881

	jobs:
	test_mi300:
	runs-on: linux-mi300-gpu-1
	container:
	image: rocm/dev-ubuntu-22.04:6.3
	options: --user root --device=/dev/kfd --device=/dev/dri --ipc=host --group-add video --cap-add=SYS_PTRACE --security-opt seccomp=unconfined

cc @yamiyysu

[stbaione]

Added the line to register container for mi300 runner

[ScottTodd]

Still missing some deps, like these:

iree/.github/workflows/pkgci_test_amd_mi300.yml

Lines 38 to 41 in e5a5881

	- name: "Install dependencies"
	run: |
	sudo apt-get update
	sudo apt-get install -y cmake ninja-build clang lld git

Trying to install those will run into #19955 though.

[ScottTodd]

Latest logs: https://github.com/iree-org/iree/actions/runs/13273266653/job/37058818527?pr=19961#step:7:29

[stbaione]

Yeah, latest run hit that error

[ScottTodd]

There is a syntax error here: https://github.com/iree-org/iree/actions/runs/13272021471

Invalid workflow file: .github/workflows/pkgci.yml#L111
The workflow is not valid. In .github/workflows/pkgci.yml (Line: 111, Col: 11): Error from called workflow iree-org/iree/.github/workflows/pkgci_test_sharktank.yml@29df51be7e68c626460b68d7e9a80f654091adb8 (Line: 25, Col: 16): Unexpected symbol: '{}'. Located at position 37 within expression: fromJSON(toJSON(matrix.container || {}))

[stbaione]

Pushed what should be a fix

[ScottTodd]

So far so good 🤞

[ScottTodd]

Please also resolve the DCO check: https://iree.dev/developers/general/contributing/#developer-certificate-of-origin