NVIDIA CMP 170HX (GA100, sm_80) 上的 GLM-5.3-Flash 推理引擎,支持 512K 上下文。
中文 | English
- AX = Axis(主线/轴心)—— 本项目在 TensorFold python-0.6 HEAD 之上建立的自研主线
- 300 = tok/s 冲刺目标
- AX300 = 以 300 tok/s 为目标的 CMP 170HX 主线引擎
本项目全部通信、量化和优化代码均为自研:
| 组件 | 出处 | 自研? |
|---|---|---|
| 通信内核(4 个 CUDA 内核) | shm_allgather.py |
✅ 100% 自研 |
| fp8-over-NCCL / wide-NB / fp8 gather | 本项目新增 | ✅ 100% 自研 |
| 基础框架 | TensorFold python-0.6 (ashhart, Apache/MIT) | 框架(类比 PyTorch) |
| 上下文 | structured | coding | prose | exactness |
|---|---|---|---|---|
| 32K | 241.6 | 181.7 | 117.1 | PASS |
| 128K | 241.2 | 181.3 | 116.7 | PASS |
| 256K | 241.5 | 181.5 | 117.3 | PASS |
| 512K | 241.1 | 181.5 | 117.0 | PASS |
| 上下文 (tok) | prefill (tok/s) | 时延 |
|---|---|---|
| 5,931 | 1,561 | 3.8s |
| 30,684 | 1,522 | 20.2s |
| 129,830 | 1,483 | 87.6s |
| 259,661 | 1,438 | 180.5s |
| 519,322 | 1,394 | 372.6s |
| 上下文 | 每卡 | 余量 |
|---|---|---|
| 32K | 46.35 GiB | 10.2 GiB |
| 512K | 55.37 GiB | 1.2 GiB |
~/tensorfold-poc/ax300_prod.sh start # ~15 分钟加载,:8000 服务
~/tensorfold-poc/ax300_prod.sh status # 健康检查
~/tensorfold-poc/ax300_prod.sh stop # 停止
# 基准(decode 三数 + exactness)
TF_BENCH_BASE=http://127.0.0.1:8000 venv/bin/python tf4_ab_bench.py --profile structured --tag <名> --exactness
# 冷 prefill(带上下文长度列)
TF_BENCH_BASE=http://127.0.0.1:8000 venv/bin/python prefill_probe.py 6000 31000 131000| 变量 | 作用 | 增益 |
|---|---|---|
TF_SHM_AG=1 |
自研 shm one-shot all-gather | +21.5% |
TF_BF16_GATHER=1 |
bf16 通信载荷 | +35.8% |
TF_FP8_GATHER=1 |
fp8 通信载荷 | +20.7% |
TF_FP8_WIDE=1 |
wide-NB 内核 | prefill 4.5× |
TF_FP8_NCCL=1 |
fp8-over-NCCL | prefill 1.7× |
TF_CONTEXT=N |
上下文(最大 524288) | — |
cmp170hx/patches/ 构建补丁(sm_80 准入等)
cmp170hx/results/ 实验收据(TEST-MATRIX.md 为总表)
cmp170hx/scripts/ 基准和测试脚本
docs/ 设计文档、PP4 规格、长上下文评估
tools/ 重合检查器等工具
- TensorFold 引擎框架:ashhart/TensorFold(Apache/MIT)
- host-shm collective 思想参考
- copy-drafts 思想:BobClawblaw(上游 PR #468)
见 LICENSE 和 THIRD_PARTY_NOTICES。
GLM-5.3-Flash inference on 4× NVIDIA CMP 170HX (sm_80, PCIe-only, no P2P). 512K context support. All communication, quantization, and optimization code is self-developed.
Performance: 241 tok/s structured / 181 coding / 117 prose decode (stable across 32K-512K context); 1,394-1,586 tok/s cold prefill. Exactness contract drafted == serial maintained.
Engine name: AX300 = "Axis 300" — the mainline branch targeting 300 tok/s on CMP 170HX.
Framework: TensorFold python-0.6 HEAD (ashhart, Apache/MIT), used as the engine framework (analogous to PyTorch for applications). Our sm_80 admission (PR #464) and GLM 4-rank (PR #465) were relanded upstream under our name.