| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
Add Sol-RL Post-Training Workflow (#369) * feat: land Sol-RL post-training workflow updates Move the post-training stack under diffusion/post_training and wire in the launch, checkpoint, dataset, and documentation updates needed to run Sol-RL directly from the Sana repo. This also surfaces Sol-RL in the main README and docs so the NVlabs branch is ready to share and use. Made-with: Cursor * refactor Sol-RL text encoding and test coverage Centralize shared prompt encoding across the Sol-RL training entrypoints so SD3, FLUX, and SANA stay aligned and easier to maintain. Clean up low-signal comments and add a minimal real-training test to catch rollout and launcher regressions earlier. Made-with: Cursor * style: apply pre-commit fixes for Sol-RL PR Sync the branch with the repository hooks so CI can clear pre-commit and unblock the PR test jobs. This commit only carries formatter, import-order, and newline normalization updates. Made-with: Cursor * update Sol-RL configs, docs, and CI * Remove Sol-RL test dependency from remove-label job - Remove test-training-sol-rl from remove-label needs - Match original repo pattern to avoid remove-label failures - Sol-RL test can fail independently without affecting label cleanup * Remove entire remove-label job from CI - Match the simpler CI structure of reference repo - Eliminates potential remove-label failures completely - Keeps CI focused on actual testing rather than label management * Update README with new SANA features and techniques * Revert LTX2VAE changes in builder.py - Restore AutoencoderKLLTX2Video import and usage - Revert commit 06af5616 changes to this file only - Keep the original LTX2VAE implementation intact * Update diffusion/model/nets/sana_multi_scale.py * Add improved remove-label job with error handling - Based on reference repo's gpu-tests.yaml implementation - Only depends on core tests (excluding Sol-RL to avoid blocking) - Uses try-catch for robust label removal - Runs on all PR outcomes to ensure cleanup * Remove dependencies from remove-label job - Remove 'needs' so label cleanup runs independently - Ensures run-tests label is always removed regardless of test outcomes - Faster label cleanup without waiting for all tests to complete * Use delayed label removal for safety - Wait for core tests to complete before removing run-tests label - Ensures tests can access label state during execution - Use always() to remove label regardless of test outcomes - Use continue-on-error to ignore remove-label success/failure - Exclude Sol-RL test to avoid blocking label cleanup * [CI-Lint] Fix code style issues with pre-commit f548e0d70802c3e2ef25865a95975f820ea94af8 --------- Co-authored-by: junsong <cjs1020440147@icloud.com> Co-authored-by: GitHub Action <action@github.com> | 6 个月前 | |
Add Sol-RL Post-Training Workflow (#369) * feat: land Sol-RL post-training workflow updates Move the post-training stack under diffusion/post_training and wire in the launch, checkpoint, dataset, and documentation updates needed to run Sol-RL directly from the Sana repo. This also surfaces Sol-RL in the main README and docs so the NVlabs branch is ready to share and use. Made-with: Cursor * refactor Sol-RL text encoding and test coverage Centralize shared prompt encoding across the Sol-RL training entrypoints so SD3, FLUX, and SANA stay aligned and easier to maintain. Clean up low-signal comments and add a minimal real-training test to catch rollout and launcher regressions earlier. Made-with: Cursor * style: apply pre-commit fixes for Sol-RL PR Sync the branch with the repository hooks so CI can clear pre-commit and unblock the PR test jobs. This commit only carries formatter, import-order, and newline normalization updates. Made-with: Cursor * update Sol-RL configs, docs, and CI * Remove Sol-RL test dependency from remove-label job - Remove test-training-sol-rl from remove-label needs - Match original repo pattern to avoid remove-label failures - Sol-RL test can fail independently without affecting label cleanup * Remove entire remove-label job from CI - Match the simpler CI structure of reference repo - Eliminates potential remove-label failures completely - Keeps CI focused on actual testing rather than label management * Update README with new SANA features and techniques * Revert LTX2VAE changes in builder.py - Restore AutoencoderKLLTX2Video import and usage - Revert commit 06af5616 changes to this file only - Keep the original LTX2VAE implementation intact * Update diffusion/model/nets/sana_multi_scale.py * Add improved remove-label job with error handling - Based on reference repo's gpu-tests.yaml implementation - Only depends on core tests (excluding Sol-RL to avoid blocking) - Uses try-catch for robust label removal - Runs on all PR outcomes to ensure cleanup * Remove dependencies from remove-label job - Remove 'needs' so label cleanup runs independently - Ensures run-tests label is always removed regardless of test outcomes - Faster label cleanup without waiting for all tests to complete * Use delayed label removal for safety - Wait for core tests to complete before removing run-tests label - Ensures tests can access label state during execution - Use always() to remove label regardless of test outcomes - Use continue-on-error to ignore remove-label success/failure - Exclude Sol-RL test to avoid blocking label cleanup * [CI-Lint] Fix code style issues with pre-commit f548e0d70802c3e2ef25865a95975f820ea94af8 --------- Co-authored-by: junsong <cjs1020440147@icloud.com> Co-authored-by: GitHub Action <action@github.com> | 6 个月前 | |
Add Sol-RL Post-Training Workflow (#369) * feat: land Sol-RL post-training workflow updates Move the post-training stack under diffusion/post_training and wire in the launch, checkpoint, dataset, and documentation updates needed to run Sol-RL directly from the Sana repo. This also surfaces Sol-RL in the main README and docs so the NVlabs branch is ready to share and use. Made-with: Cursor * refactor Sol-RL text encoding and test coverage Centralize shared prompt encoding across the Sol-RL training entrypoints so SD3, FLUX, and SANA stay aligned and easier to maintain. Clean up low-signal comments and add a minimal real-training test to catch rollout and launcher regressions earlier. Made-with: Cursor * style: apply pre-commit fixes for Sol-RL PR Sync the branch with the repository hooks so CI can clear pre-commit and unblock the PR test jobs. This commit only carries formatter, import-order, and newline normalization updates. Made-with: Cursor * update Sol-RL configs, docs, and CI * Remove Sol-RL test dependency from remove-label job - Remove test-training-sol-rl from remove-label needs - Match original repo pattern to avoid remove-label failures - Sol-RL test can fail independently without affecting label cleanup * Remove entire remove-label job from CI - Match the simpler CI structure of reference repo - Eliminates potential remove-label failures completely - Keeps CI focused on actual testing rather than label management * Update README with new SANA features and techniques * Revert LTX2VAE changes in builder.py - Restore AutoencoderKLLTX2Video import and usage - Revert commit 06af5616 changes to this file only - Keep the original LTX2VAE implementation intact * Update diffusion/model/nets/sana_multi_scale.py * Add improved remove-label job with error handling - Based on reference repo's gpu-tests.yaml implementation - Only depends on core tests (excluding Sol-RL to avoid blocking) - Uses try-catch for robust label removal - Runs on all PR outcomes to ensure cleanup * Remove dependencies from remove-label job - Remove 'needs' so label cleanup runs independently - Ensures run-tests label is always removed regardless of test outcomes - Faster label cleanup without waiting for all tests to complete * Use delayed label removal for safety - Wait for core tests to complete before removing run-tests label - Ensures tests can access label state during execution - Use always() to remove label regardless of test outcomes - Use continue-on-error to ignore remove-label success/failure - Exclude Sol-RL test to avoid blocking label cleanup * [CI-Lint] Fix code style issues with pre-commit f548e0d70802c3e2ef25865a95975f820ea94af8 --------- Co-authored-by: junsong <cjs1020440147@icloud.com> Co-authored-by: GitHub Action <action@github.com> | 6 个月前 | |
Add Sol-RL Post-Training Workflow (#369) * feat: land Sol-RL post-training workflow updates Move the post-training stack under diffusion/post_training and wire in the launch, checkpoint, dataset, and documentation updates needed to run Sol-RL directly from the Sana repo. This also surfaces Sol-RL in the main README and docs so the NVlabs branch is ready to share and use. Made-with: Cursor * refactor Sol-RL text encoding and test coverage Centralize shared prompt encoding across the Sol-RL training entrypoints so SD3, FLUX, and SANA stay aligned and easier to maintain. Clean up low-signal comments and add a minimal real-training test to catch rollout and launcher regressions earlier. Made-with: Cursor * style: apply pre-commit fixes for Sol-RL PR Sync the branch with the repository hooks so CI can clear pre-commit and unblock the PR test jobs. This commit only carries formatter, import-order, and newline normalization updates. Made-with: Cursor * update Sol-RL configs, docs, and CI * Remove Sol-RL test dependency from remove-label job - Remove test-training-sol-rl from remove-label needs - Match original repo pattern to avoid remove-label failures - Sol-RL test can fail independently without affecting label cleanup * Remove entire remove-label job from CI - Match the simpler CI structure of reference repo - Eliminates potential remove-label failures completely - Keeps CI focused on actual testing rather than label management * Update README with new SANA features and techniques * Revert LTX2VAE changes in builder.py - Restore AutoencoderKLLTX2Video import and usage - Revert commit 06af5616 changes to this file only - Keep the original LTX2VAE implementation intact * Update diffusion/model/nets/sana_multi_scale.py * Add improved remove-label job with error handling - Based on reference repo's gpu-tests.yaml implementation - Only depends on core tests (excluding Sol-RL to avoid blocking) - Uses try-catch for robust label removal - Runs on all PR outcomes to ensure cleanup * Remove dependencies from remove-label job - Remove 'needs' so label cleanup runs independently - Ensures run-tests label is always removed regardless of test outcomes - Faster label cleanup without waiting for all tests to complete * Use delayed label removal for safety - Wait for core tests to complete before removing run-tests label - Ensures tests can access label state during execution - Use always() to remove label regardless of test outcomes - Use continue-on-error to ignore remove-label success/failure - Exclude Sol-RL test to avoid blocking label cleanup * [CI-Lint] Fix code style issues with pre-commit f548e0d70802c3e2ef25865a95975f820ea94af8 --------- Co-authored-by: junsong <cjs1020440147@icloud.com> Co-authored-by: GitHub Action <action@github.com> | 6 个月前 | |
Add Sol-RL Post-Training Workflow (#369) * feat: land Sol-RL post-training workflow updates Move the post-training stack under diffusion/post_training and wire in the launch, checkpoint, dataset, and documentation updates needed to run Sol-RL directly from the Sana repo. This also surfaces Sol-RL in the main README and docs so the NVlabs branch is ready to share and use. Made-with: Cursor * refactor Sol-RL text encoding and test coverage Centralize shared prompt encoding across the Sol-RL training entrypoints so SD3, FLUX, and SANA stay aligned and easier to maintain. Clean up low-signal comments and add a minimal real-training test to catch rollout and launcher regressions earlier. Made-with: Cursor * style: apply pre-commit fixes for Sol-RL PR Sync the branch with the repository hooks so CI can clear pre-commit and unblock the PR test jobs. This commit only carries formatter, import-order, and newline normalization updates. Made-with: Cursor * update Sol-RL configs, docs, and CI * Remove Sol-RL test dependency from remove-label job - Remove test-training-sol-rl from remove-label needs - Match original repo pattern to avoid remove-label failures - Sol-RL test can fail independently without affecting label cleanup * Remove entire remove-label job from CI - Match the simpler CI structure of reference repo - Eliminates potential remove-label failures completely - Keeps CI focused on actual testing rather than label management * Update README with new SANA features and techniques * Revert LTX2VAE changes in builder.py - Restore AutoencoderKLLTX2Video import and usage - Revert commit 06af5616 changes to this file only - Keep the original LTX2VAE implementation intact * Update diffusion/model/nets/sana_multi_scale.py * Add improved remove-label job with error handling - Based on reference repo's gpu-tests.yaml implementation - Only depends on core tests (excluding Sol-RL to avoid blocking) - Uses try-catch for robust label removal - Runs on all PR outcomes to ensure cleanup * Remove dependencies from remove-label job - Remove 'needs' so label cleanup runs independently - Ensures run-tests label is always removed regardless of test outcomes - Faster label cleanup without waiting for all tests to complete * Use delayed label removal for safety - Wait for core tests to complete before removing run-tests label - Ensures tests can access label state during execution - Use always() to remove label regardless of test outcomes - Use continue-on-error to ignore remove-label success/failure - Exclude Sol-RL test to avoid blocking label cleanup * [CI-Lint] Fix code style issues with pre-commit f548e0d70802c3e2ef25865a95975f820ea94af8 --------- Co-authored-by: junsong <cjs1020440147@icloud.com> Co-authored-by: GitHub Action <action@github.com> | 6 个月前 | |
Add Sol-RL Post-Training Workflow (#369) * feat: land Sol-RL post-training workflow updates Move the post-training stack under diffusion/post_training and wire in the launch, checkpoint, dataset, and documentation updates needed to run Sol-RL directly from the Sana repo. This also surfaces Sol-RL in the main README and docs so the NVlabs branch is ready to share and use. Made-with: Cursor * refactor Sol-RL text encoding and test coverage Centralize shared prompt encoding across the Sol-RL training entrypoints so SD3, FLUX, and SANA stay aligned and easier to maintain. Clean up low-signal comments and add a minimal real-training test to catch rollout and launcher regressions earlier. Made-with: Cursor * style: apply pre-commit fixes for Sol-RL PR Sync the branch with the repository hooks so CI can clear pre-commit and unblock the PR test jobs. This commit only carries formatter, import-order, and newline normalization updates. Made-with: Cursor * update Sol-RL configs, docs, and CI * Remove Sol-RL test dependency from remove-label job - Remove test-training-sol-rl from remove-label needs - Match original repo pattern to avoid remove-label failures - Sol-RL test can fail independently without affecting label cleanup * Remove entire remove-label job from CI - Match the simpler CI structure of reference repo - Eliminates potential remove-label failures completely - Keeps CI focused on actual testing rather than label management * Update README with new SANA features and techniques * Revert LTX2VAE changes in builder.py - Restore AutoencoderKLLTX2Video import and usage - Revert commit 06af5616 changes to this file only - Keep the original LTX2VAE implementation intact * Update diffusion/model/nets/sana_multi_scale.py * Add improved remove-label job with error handling - Based on reference repo's gpu-tests.yaml implementation - Only depends on core tests (excluding Sol-RL to avoid blocking) - Uses try-catch for robust label removal - Runs on all PR outcomes to ensure cleanup * Remove dependencies from remove-label job - Remove 'needs' so label cleanup runs independently - Ensures run-tests label is always removed regardless of test outcomes - Faster label cleanup without waiting for all tests to complete * Use delayed label removal for safety - Wait for core tests to complete before removing run-tests label - Ensures tests can access label state during execution - Use always() to remove label regardless of test outcomes - Use continue-on-error to ignore remove-label success/failure - Exclude Sol-RL test to avoid blocking label cleanup * [CI-Lint] Fix code style issues with pre-commit f548e0d70802c3e2ef25865a95975f820ea94af8 --------- Co-authored-by: junsong <cjs1020440147@icloud.com> Co-authored-by: GitHub Action <action@github.com> | 6 个月前 | |
Add Sol-RL Post-Training Workflow (#369) * feat: land Sol-RL post-training workflow updates Move the post-training stack under diffusion/post_training and wire in the launch, checkpoint, dataset, and documentation updates needed to run Sol-RL directly from the Sana repo. This also surfaces Sol-RL in the main README and docs so the NVlabs branch is ready to share and use. Made-with: Cursor * refactor Sol-RL text encoding and test coverage Centralize shared prompt encoding across the Sol-RL training entrypoints so SD3, FLUX, and SANA stay aligned and easier to maintain. Clean up low-signal comments and add a minimal real-training test to catch rollout and launcher regressions earlier. Made-with: Cursor * style: apply pre-commit fixes for Sol-RL PR Sync the branch with the repository hooks so CI can clear pre-commit and unblock the PR test jobs. This commit only carries formatter, import-order, and newline normalization updates. Made-with: Cursor * update Sol-RL configs, docs, and CI * Remove Sol-RL test dependency from remove-label job - Remove test-training-sol-rl from remove-label needs - Match original repo pattern to avoid remove-label failures - Sol-RL test can fail independently without affecting label cleanup * Remove entire remove-label job from CI - Match the simpler CI structure of reference repo - Eliminates potential remove-label failures completely - Keeps CI focused on actual testing rather than label management * Update README with new SANA features and techniques * Revert LTX2VAE changes in builder.py - Restore AutoencoderKLLTX2Video import and usage - Revert commit 06af5616 changes to this file only - Keep the original LTX2VAE implementation intact * Update diffusion/model/nets/sana_multi_scale.py * Add improved remove-label job with error handling - Based on reference repo's gpu-tests.yaml implementation - Only depends on core tests (excluding Sol-RL to avoid blocking) - Uses try-catch for robust label removal - Runs on all PR outcomes to ensure cleanup * Remove dependencies from remove-label job - Remove 'needs' so label cleanup runs independently - Ensures run-tests label is always removed regardless of test outcomes - Faster label cleanup without waiting for all tests to complete * Use delayed label removal for safety - Wait for core tests to complete before removing run-tests label - Ensures tests can access label state during execution - Use always() to remove label regardless of test outcomes - Use continue-on-error to ignore remove-label success/failure - Exclude Sol-RL test to avoid blocking label cleanup * [CI-Lint] Fix code style issues with pre-commit f548e0d70802c3e2ef25865a95975f820ea94af8 --------- Co-authored-by: junsong <cjs1020440147@icloud.com> Co-authored-by: GitHub Action <action@github.com> | 6 个月前 |
| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
| 6 个月前 | ||
| 6 个月前 | ||
| 6 个月前 | ||
| 6 个月前 | ||
| 6 个月前 | ||
| 6 个月前 | ||
| 6 个月前 |