refactor: reusable workflows accept config-file input (PoC)

[Eren-Jeager123]

Summary

• Add optional config-file input to 4 PyTorch-relevant reusable workflows (sanity, security, telemetry, EFA)
• When config-file is provided, the reusable workflow reads metadata from the YAML config directly — callers no longer need to parse and forward 10+ individual parameters
• Backward compatible: existing callers (vLLM, SGLang, Ray, Base) pass individual inputs and work unchanged
• Updated pr-pytorch-ec2-cuda.yml as PoC — eliminates the verbose load-config→parse→forward pattern

Motivation

GitHub Actions limitation: when a matrix job has outputs, only the last-finishing matrix leg gets its outputs exposed. If multiple PyTorch versions change in the same PR, downstream reusable workflow calls only receive metadata from one version.

By making reusable workflows self-sufficient (reading their own config), each matrix leg can pass its own config-file path without shared outputs clobbering each other.

Workflow & Caller Inventory

13 Reusable Workflows

#	Reusable Workflow	Key Inputs
1	reusable-sanity-tests.yml	image-uri, framework, framework-version, python-version, cuda-version, os-version, customer-type, arch-type, device-type, contributor, container-type
2	reusable-security-tests.yml	image-uri, framework, framework-version
3	reusable-telemetry-tests.yml	image-uri, framework, framework-version, container-type
4	reusable-efa-tests.yml	image-uri, aws-account-id, aws-region
5	reusable-vllm-model-tests.yml	image-uri, aws-account-id, aws-region
6	reusable-vllm-upstream-tests.yml	image-uri, framework-version, vllm-ref, setup-script, example-test-script
7	reusable-vllm-sagemaker-tests.yml	image-uri only
8	reusable-vllm-omni-model-tests.yml	image-uri, customer-type
9	reusable-sglang-model-tests.yml	image-uri, aws-account-id, aws-region
10	reusable-sglang-upstream-tests.yml	image-uri, framework-version, benchmark-start-command
11	reusable-sglang-sagemaker-tests.yml	image-uri only
12	reusable-release-image.yml	source-image-uri, release-spec, environment
13	reusable-sagemaker-xgboost-integ-tests.yml	image-uri, aws-account-id, aws-region

PR Workflows (20 callers)

Caller	Reusable(s) Called	Multi-version Matrix?
pr-pytorch-ec2-cuda.yml (this PoC)	sanity, security, telemetry, efa	Future: yes
pr-pytorch-ec2-cpu.yml	sanity, security, telemetry	Future: yes
pr-pytorch-sagemaker-cpu.yml	sanity, security, telemetry	Future: yes
pr-pytorch-sagemaker-cuda.yml	sanity, security, telemetry	Future: yes
pr-base-v1.yml	sanity, security	No
pr-base-v2.yml	sanity, security	No (matrix over targets)
pr-vllm-ec2-amzn2023.yml	sanity, security, telemetry, upstream, model	No
pr-vllm-ec2.yml	sanity, security, telemetry, upstream, model	No
pr-vllm-sagemaker-amzn2023.yml	sanity, security, telemetry, sagemaker	No
pr-vllm-sagemaker.yml	sanity, security, telemetry, sagemaker	No
pr-vllm-hyperpod-amzn2023.yml	sanity, security, telemetry	No
pr-vllm-omni-ec2-amzn2023.yml	sanity, security, telemetry, vllm-omni-model	No
pr-vllm-omni-sagemaker-amzn2023.yml	sanity, security, telemetry, vllm-omni-model	No
pr-sglang-ec2-amzn2023.yml	sanity, security, telemetry, sglang-upstream, sglang-model	No
pr-sglang-ec2.yml	sanity, security, telemetry, sglang-upstream, sglang-model	No
pr-sglang-sagemaker-amzn2023.yml	sanity, security, telemetry, sglang-sagemaker	No
pr-sglang-sagemaker.yml	sanity, security, telemetry, sglang-sagemaker	No
pr-ray-ec2-cpu.yml	sanity, security	No
pr-ray-ec2-gpu.yml	sanity, security	No
pr-ray-sagemaker-cpu/gpu.yml	sanity, security	No
pr-sagemaker-xgboost.yml	xgboost-integ	No

Autorelease Workflows (14 callers)

Caller	Reusable(s) Called
autorelease-pytorch-ec2-cuda.yml	sanity, security, telemetry, efa, release
autorelease-pytorch-ec2-cpu.yml	sanity, security, telemetry, release
autorelease-pytorch-sagemaker-cpu.yml	sanity, security, telemetry, release
autorelease-pytorch-sagemaker-cuda.yml	sanity, security, telemetry, efa, release
autorelease-vllm-ec2-amzn2023.yml	sanity, security, telemetry, upstream, model, release
autorelease-vllm-ec2.yml	sanity, security, telemetry, upstream, model, release
autorelease-vllm-sagemaker-amzn2023.yml	sanity, security, telemetry, sagemaker, release
autorelease-vllm-sagemaker.yml	sanity, security, telemetry, sagemaker, release
autorelease-vllm-hyperpod-amzn2023.yml	sanity, security, telemetry, release
autorelease-vllm-omni.yml	sanity, security, telemetry, vllm-omni-model, release
autorelease-sglang-ec2.yml	sanity, security, telemetry, sglang-upstream, sglang-model, release
autorelease-sglang-sagemaker.yml	sanity, security, telemetry, sglang-sagemaker, release
autorelease-ray.yml	sanity, security, release
autorelease-base.yml	sanity, security, release

Dispatch Workflows (3 callers)

Caller	Reusable(s) Called
dispatch-partner-release.yml	release
dispatch-release-sagemaker-xgboost.yml	xgboost-integ, release

Design

Before (10+ inputs forwarded per call):
┌─────────────────┐     ┌──────────────────────────────────────┐
│ PR Workflow      │     │ Reusable Workflow                    │
│                  │     │                                      │
│ load-config ────────┐  │  inputs:                            │
│   parse YAML    │   │  │    framework: "pytorch"             │
│   forward each  │   ├──│    framework-version: "2.11.0"      │
│   field as      │   │  │    python-version: "py312"          │
│   input         │   │  │    cuda-version: "cu130"            │
│                  │   │  │    os-version: "amzn2023"           │
│                  │   │  │    ... (10+ more)                   │
└─────────────────┘   │  └──────────────────────────────────────┘
                      └──→  (each field passed individually)

After (single config-file input):
┌─────────────────┐     ┌──────────────────────────────────────┐
│ PR Workflow      │     │ Reusable Workflow                    │
│                  │     │                                      │
│ config-file ─────────── │  inputs:                            │
│   just pass     │     │    config-file: "path/to/config.yml" │
│   the path      │     │                                      │
│                  │     │  load-config job:                    │
│                  │     │    reads YAML → resolves all fields  │
└─────────────────┘     └──────────────────────────────────────┘

Blast Radius

• Only pr-pytorch-ec2-cuda.yml is triggered by this PR (it watches its own path)
• Non-pytorch workflows do NOT have reusable-sanity/security/telemetry/efa in their path triggers
• Backward compatible: vLLM/SGLang/Ray/Base callers pass individual inputs → else branch forwards them unchanged

Test plan

• Verify YAML syntax passes (done locally)
• Confirm only PR - PyTorch EC2 CUDA workflow triggers on this PR
• Validate that existing callers (vLLM, SGLang, etc.) work unchanged on their next PR
• Future: migrate remaining 3 PyTorch PR workflows + 4 autorelease workflows to config-file interface