[NPU] Sync Ascend with main and consolidate NPU adaptations

[Meihan-chen]

Summary

Bring Ascend up to date with main through ce92eff1, including the new native
model-loading, weight-transfer and accelerator interfaces, while incorporating
Ascend's training-stack updates and GLM-4.7-Flash support through 083aea63.

The integration follows main's shared training and rollout flow. Necessary NPU
differences are kept in small platform/provider interfaces and
docker/npu_patch, rather than maintaining a separate copy of main's
orchestration. This keeps subsequent main merges easier to maintain.

Main changes and corresponding NPU adaptations

Area	Changes from main	NPU adaptation in this PR
Weight synchronization	Native stateful trainer/worker weight-transfer interfaces replace the legacy worker-extension path; colocated transfer uses packed IPC payloads.	Reuse the shared update sequence, bucket packing and per-engine-slot coordination. Select hccl / npu_ipc and their initialization types through the NPU provider; implement the matching native worker contracts in vendor patches and remove obsolete Vime worker extensions.
MoE weight transfer	Rank-local expert routing and payloads avoid treating every rank as a full-model weight source.	Support the corresponding HCCL/IPC payloads, list workers and empty payload boundaries; preserve MoE loader metadata and routing/control buffers during reload.
Model loading and Bridge	Native HF-to-Megatron loading and Megatron-to-HF export replace the Bridge-based model integration and legacy conversion-mode arguments.	Follow the native loader/exporter structure for Qwen3-VL and GLM-4.7-Flash, including multimodal configuration, weight-prefix mapping and MLA/MoE/MTP mapping coverage. Do not reintroduce Bridge model plugins.
Accelerator and rollout structure	Shared accelerator APIs and the rollout engine_group.py split centralize device operations and engine management.	Register the NPU accelerator, order training bootstrap correctly, and carry NPU resource/visible-device differences into the new engine-group structure without duplicating rollout orchestration.
Training/serving memory lifecycle	Shared training contexts and serving sleep/wake coordinate memory reuse between training and generation.	Initialize NPU optimizer state inside the existing training memory context; fix packed IPC reference ownership and reclamation; keep serving block-table/control tensors outside releasable KV-cache storage.
MTP updates	The shared flow updates the main model and draft model through separate weight-update transactions.	Add the Ascend worker's draft-target lifecycle and GLM MTP graph-compatible position mask, keeping the common Vime update flow unchanged.

Native loading here refers to the model construction/loading/export path.
The NPU image still contains Bridge-related dependencies; removing all of those
packages is not part of this PR.

Additional Ascend integration

• Integrate standalone MegatronAdaptor and TransformerEngineNPU, with the common Megatron patch applied before the NPU-specific patch.
• Rebase NPU vendor patches onto the paired serving sources: vLLM e6bfe03a and vLLM-Ascend fd815467.
• Make torch_dist conversion/reference loading the default Qwen3-30B-A3B test path; retain HF-only testing via VIME_TEST_TORCH_DIST_REF_LOAD=0.
• Add GLM-4.7-Flash non-colocate coverage. MTP is off by default and explicitly enabled with VIME_TEST_GLM_MTP=1; VIME_TEST_GLM_EAGER=1 independently enables eager diagnosis.
• Update existing mapping, platform and launch-contract tests, remove obsolete CLI arguments, and keep local model/dataset/checkpoint overrides out of the commits.

Validation

Local E2E results collected during integration:

Model / mode	Coverage	Result
Qwen3-4B	Non-colocate stateful HCCL weight updates, training and rollout	PASS
Qwen3-30B-A3B, HF	Colocate packed NPU IPC, rank-local MoE, training and rollout	PASS
Qwen3-30B-A3B, torch_dist	Checkpoint conversion, reference loading, training and rollout	PASS
Qwen3-VL-8B	Native loading, training TP4/SP, rollout TP1 per slot	PASS
GLM-4.7-Flash, no MTP	Native loading and non-colocate HCCL	PASS
GLM-4.7-Flash, MTP eager	Two training steps, main/draft updates and subsequent generation	PASS
GLM-4.7-Flash, MTP graph	Actual graph capture, two training steps and main/draft updates	PASS

Not yet adapted / follow-ups

• Qwen3.5 NPU: the tested environment has incompatible serving/training GDN operator stacks: overlapping FLA/serving operator libraries and a serving causal-convolution Python-to-binary schema mismatch. It needs a matched serving binary/OPP and training FLA environment before completing NPU adaptation and E2E validation. Main's model implementation remains available.
• GLM colocate, other unvalidated model/MTP combinations, Omni, DSpark, speculative metrics and pull/delta weight transfer are outside the validated NPU scope.
• Broader parallel configurations, long-running memory stability and performance equivalence still need dedicated validation.

[gemini-code-assist]

Code Review

This pull request introduces significant updates to the vime framework, including the addition of a delta weight synchronization mechanism for training/inference disaggregation, support for agentic RL workflows with new sandbox backends, and improvements to the CI/CD pipeline using Buildkite. My review identified two critical issues: a logic error in the pull_checkpoint version rollback condition and a token duplication/length mismatch error in the MemAgent rollout logic. Both issues require immediate attention to ensure correct checkpoint management and training stability.

[gemini-code-assist]

In pull_checkpoint, if target_version < applied (e.g., during a rollback or reset to an older checkpoint), start (which is target_version) will be less than applied. Since start > applied is False, the checkpoint reset logic will be skipped, and the local checkpoint will remain at the newer applied version instead of rolling back to the requested target_version. Changing the condition to start != applied ensures that both forward-resets and backward-rollbacks are correctly handled.

        if applied is None or start != applied:

[gemini-code-assist]

In generate, cat_token_ids already starts with the prompt of the first turn (first_out.prompt_token_ids) because turns[0]['tokens'] is constructed as out.prompt_token_ids + out.token_ids. Prepending prompt_token_ids to sample.tokens here duplicates the first prompt. Furthermore, since sample.loss_mask and sample.rollout_log_probs are set directly to cat_loss_mask and cat_log_probs (which do not have prompt_token_ids prepended), this causes a length mismatch between tokens, loss_mask, and log_probs, which will crash the training process. Setting sample.tokens = cat_token_ids resolves both the duplication and the length mismatch.

Suggested change

	sample.tokens = prompt_token_ids + cat_token_ids
	sample.tokens = cat_token_ids

[read-the-docs-community]

Documentation build overview

📚 vime | 🛠️ Build #34490796 | 📁 Comparing edf744e against latest (ce92eff)

  🔍 Preview build  

48 files changed · + 3 added · ± 45 modified

+ Added

• get_started/NPU.html
• _examples_synced/retool/README.html
• _examples_synced/search-r1/README.html

± Modified

• genindex.html
• index.html
• search.html
• advanced/arch-support-beyond-megatron.html
• advanced/delta-weight-sync.html
• advanced/external-rollout-engines.html
• advanced/fault-tolerance.html
• advanced/low-precision.html
• advanced/megatron-config.html
• advanced/observability.html
• and 35 more...

[momo609]

amd patch can remove

[momo609]

The GDN Triton operator has been removed. Is that because Qwen3.5’s training side is already supported in another third-party repo?

[Meihan-chen]

qwen3-4B non-colocate