feat: Support vLLM data parallel calibration for hy_v3

[jds250]

Overview

This PR adds Ray actor-managed data parallelism to the vLLM-based calibration workflow and integrates it into the Hy3 FP8 post-training quantization pipeline.

Motivation

Calibrating large Hy3 models with a single vLLM instance can be time-consuming and may not fully utilize available multi-node and multi-GPU resources.

This PR enables calibration samples to be sharded across multiple data-parallel replicas. Each DP replica can still use tensor parallelism internally through vLLM. After calibration, the driver merges the activation, MoE, MTP, and KV-cache statistics collected by all replicas into the same artifacts expected by the existing quantization stage.

In addition, the calibration and quantization stages can now share a single YAML configuration, reducing duplicated command-line arguments and making the overall PTQ workflow easier to configure, reproduce, and maintain.

Key Changes

tools/run_vllm_calibrate.py

• Added data-parallel calibration options, including:

  • dp_size

• Refactored the calibration logic into run_one_calibration, allowing the same implementation to run in:

  • single-replica mode
  • Ray actor-managed DP mode

• Sharded calibration prompts and KV scale search prompts according to dp_rank.

• Added per-rank result payloads for cross-DP aggregation.

• Added support for a second-stage KV-cache scale search after merged activation statistics become available.

• Preserved support for:

  • per-tensor KV-cache calibration
  • per-head KV-cache calibration
  • MTP activation statistics
  • MTP MoE statistics

angelslim/utils/vllm_calibration_dp.py

• Added a Ray actor-managed launcher for data-parallel calibration.

• Creates one placement group and one long-lived actor for each DP rank.

• Validates DP/TP resource requirements and placement-group scheduling feasibility.

• Keeps each actor's vLLM instance alive so that the follow-up KV scale search can reuse the loaded model.

• Merges DP-local:

  • activation statistics
  • MoE expert statistics
  • MTP activation statistics
  • MTP MoE statistics
  • KV-cache statistics

• Aggregates KV quantization error profiles across DP replicas and writes the final merged outputs:

  • kv_scale_multipliers*.json
  • kv_cache_tuned_scales*.json

angelslim/utils/__init__.py

• Exported the newly added vLLM data-parallel calibration helper APIs.

angelslim/compressor/quant/core/vllm_calibrate_utils/search.py

• Added DP-friendly KV-cache MSE profile collection.

• Supports both:

  • per-tensor KV scale search
  • per-head KV scale search

• Preserves the existing local search logic while enabling cross-DP aggregation through additive MSE/SSE profiles.

angelslim/compressor/quant/core/vllm_calibrate_utils/__init__.py

• Exported the newly added KV profile collectors and result-processing helpers.

configs/Hy3/ptq/fp8/Hy3_vllm_ptq_per_tensor.yaml

• Added a configuration for the Hy3 FP8 PTQ workflow.
  • DP settings——dp_size

Output Artifacts

The updated pipeline preserves the existing stage-1 output names and generates DP-aggregated KV scale-search outputs when enabled:

activation_stats.json
moe_expert_stats.json
mtp_activation_stats.json
mtp_moe_expert_stats.json
kv_scale_multipliers.json
kv_cache_tuned_scales.json
kv_scale_multipliers_per_head.json
kv_cache_tuned_scales_per_head.json

Summary

This PR introduces the following improvements:

1. Scales Hy3 vLLM calibration across multiple data-parallel replicas.
2. Merges activation, MoE, MTP, and KV-cache statistics across DP replicas.
3. Aggregates KV-cache scale-search error profiles across DP replicas.
4. Preserves compatibility with the existing FP8 quantization workflow and artifact formats.

[yghstill]

@jds250
Please per-commit code formatting:

pip3 install pre-commit black isort flake8
cd AngelSlim
pre-commit install

[WOODchen7]

这个import放前面

[WOODchen7]

还有这个import