feat(quant): add triton fp8 w8a8 per-tensor quantization method

[sufubao]

What

Add triton-fp8w8a8-pertensor (aliases fp8w8a8-pertensor, triton-fp8w8a8-pt, fp8w8a8-pt) — an FP8 w8a8 quantization method with no vLLM dependency: weight is per-tensor FP8, activation is per-token FP8, both via Triton kernels. The fp8 GEMM dispatches to the best available backend.

新增 triton-fp8w8a8-pertensor（别名 fp8w8a8-pertensor / triton-fp8w8a8-pt / fp8w8a8-pt）：不依赖 vLLM 的 FP8 w8a8 量化方法，权重 per-tensor、激活 per-token，全部走 Triton kernel；fp8 GEMM 自动选择最优后端。

Why

The existing vllm-fp8w8a8* methods require vLLM. On images without vLLM (or when a pure-Triton path is preferred) there was no fp8 w8a8 option. This fills that gap and reaches vLLM-comparable end-to-end fp8 throughput by routing the GEMM to a Cutlass-class kernel when available.

现有 vllm-fp8w8a8* 强依赖 vLLM；无 vLLM 的镜像（或希望走纯 Triton 路径）下缺少 fp8 w8a8 方案。本方法补上该能力，并在存在 Cutlass 级 kernel 时把 GEMM 路由过去，端到端吞吐与 vLLM 持平。

Changes

• lightllm/common/quantization/w8a8.py
  • New TritonFP8w8a8PerTensorQuantizationMethod: per-tensor weight quant (_fp8_per_tensor_quant), per-token activation quant via the pure-Triton lightllm_per_token_group_quant_fp8 (not the wrapper, which would dispatch to the sgl kernel that rejects group_size == k).
  • Deferred staging for fused multi-split weights and per-expert MoE weights (_try_finalize_deferred_weight).
  • fp8 GEMM backend selector (LIGHTLLM_FP8_GEMM).
• lightllm/common/basemodel/triton_kernel/quantization/scaled_mm_per_token_kernel.py: B_SCALE_IS_TENSOR path so the per-token scaled_mm accepts a per-tensor weight scale [1], not only per-channel [N].
• lightllm/common/basemodel/layer_infer/fused_moe/grouped_fused_moe.py: WEIGHT_SCALE_PER_TENSOR support in the fused MoE matmul.
• lightllm/server/api_cli.py: list triton-fp8w8a8-pertensor in the --quant_method help.

Usage

--quant_method triton-fp8w8a8-pertensor
# optional GEMM backend, default auto
export LIGHTLLM_FP8_GEMM=auto   # auto | cutlass | sgl | triton

auto picks cutlass (vLLM cutlass_scaled_mm) → sgl (sgl_kernel.fp8_scaled_mm) → triton, by availability, so images without a Cutlass-class kernel fall back to Triton cleanly. sgl_kernel rejects a per-tensor weight scale, so the scalar is expanded to per-channel [N] once and cached on the weight_pack.

Validation (H200, Qwen3.5-27B)

• End-to-end fp8 throughput ≈ 1.19x over bf16 (matches vLLM ~1.20x; the pure-triton GEMM alone capped at ~1.10x because its grid underfills the GPU on the attention-projection shapes at decode batch sizes).
• GSM8K accuracy by backend: cutlass 0.850 / sgl 0.860 / triton 0.825 / bf16 0.860 (sgl matches bf16; cutlass/triton within ~0.01–0.035).
• Correctness: per-token scale exact vs max(|x|)/448; matmul cosine 0.9993 vs bf16 ref (above the repo's 0.99 gate) across M = 1/16/47/128/7 and non-aligned K/N.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.