[Feature] Add V100 (SM70) GPU support#1

Open

mattheliu wants to merge 65 commits into

Commits on Feb 9, 2026

Commits on Feb 10, 2026

Commits on Feb 11, 2026

[BugFix]fix console log metrics waitting queue count (#6432)
CSWYF3634076
authored
[CI] 【Hackathon 10th Spring No.22】功能模块 fastdeploy/cache_manager/cache_transfer_manager.py 单测补充 (#6157)

kesmeey
and
CSWYF3634076
authored
[Feature] Add V100 (SM70) GPU support

mattheliu
and
claude
committed
[Fix] Skip RoPE in V100FlashAttentionBackend to avoid shape mismatch

mattheliu
and
claude
committed
[Fix] V100 fallback to NATIVE_ATTN instead of V100_FLASH_ATTN

mattheliu
and
claude
committed
[Fix] Handle qkv tensor splitting in PaddleNativeAttnBackend

mattheliu
and
claude
committed
[Fix] V100FlashAttentionBackend use scaled_dot_product_attention instead of flash_attn_unpadded

mattheliu
and
claude
committed
[Restore] Revert native_paddle_backend.py to upstream official version

mattheliu
and
claude
committed
[Fix] Handle dummy/profile run in V100FlashAttentionBackend

mattheliu
and
claude
committed
[Fix] Use manual attention to avoid Paddle's GQA validation in V100FlashAttentionBackend

mattheliu
and
claude
committed
[Fix] Fix batch_id indexing in forward_mixed for V100FlashAttentionBackend

mattheliu
and
claude
committed
[Fix] Use seq_lens_this_time.shape[0] as batch_size for consistency

mattheliu
and
claude
committed
[Fix] Add RoPE (Rotary Position Embedding) support to V100FlashAttentionBackend

mattheliu
and
claude
committed
[Fix] Cast RoPE output back to original dtype (float16)

mattheliu
and
claude
committed
[Fix] Use float32 for attention score computation for numerical stability

mattheliu
and
claude
committed
[Perf] Vectorize RoPE application to avoid slow per-token loop

mattheliu
and
claude
committed
[Debug] Add RoPE debug logging and fix interleaved style support

mattheliu
and
claude
committed
[Fix] Fix interleaved RoPE shape mismatch - rotary_embs already head_dim//2

mattheliu
and
claude
committed
[Debug] Add more decode phase position debugging

mattheliu
and
claude
committed
[Fix] Correct position calculation in RoPE and KV cache write

mattheliu
and
claude
committed
[Fix] Distinguish prefill vs decode for correct position calculation

mattheliu
and
claude
committed
[Fix] Fix total_seq_lens calculation for KV cache reading in V100 backend

mattheliu
and
claude
committed
[Debug] Add more detailed debug logging for V100 RoPE position calculation

mattheliu
and
claude
committed
[Debug] Add decode phase debug logging

mattheliu
and
claude
committed
[Fix] Auto-disable CUDA graph for V100 (SM70) compatibility

mattheliu
and
claude
committed
[Fix] Fix neox-style RoPE in V100 backend for Qwen3 compatibility

mattheliu
and
claude
committed