Gahow Wang 9f1fbbb98b quantization: add FP8 E4M3 W8A16 for gpt-oss MoE expert weights ...

Store expert gate_up_proj and down_proj weights in FP8 E4M3 (1 byte/elem)
with per-expert FP32 scale factors. At inference, a fused CUDA kernel
dequantizes to BF16 before the existing cuBLAS batched GEMM.

Results on gpt-oss-20b (50-problem GSM8K subset):
  - FP8 TP=1: 47/50 = 94.0% (single RTX 5090, ~25 GB VRAM)
  - BF16 TP=2: 47/50 = 94.0% (requires 2× RTX 5090, ~39 GB total)

No measurable accuracy degradation. Model size: 41.8 GB → 22.7 GB (−46%).

New files:
  - tools/quantize_fp8.py: offline BF16→FP8 conversion script
  - csrc/quantization/dequant_fp8.cu: per-expert-scale dequant kernel
  - crates/xserv-kernels/src/quantization.rs: Rust FFI wrapper
  - tools/eval_gsm8k_batch.sh: GSM8K accuracy evaluation harness

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

2026-06-07 19:33:07 +08:00

..

bench

bench: PP harness (xserv --pp vs llama.cpp -sm layer)

2026-05-29 18:45:59 +08:00

__init__.py

bench: run one server at a time, match thinking mode, fix tools package

2026-05-28 11:40:07 +08:00

analyze_divergence.py

phase 9: KV cache + autoregressive generation

2026-05-21 23:39:41 +08:00

bench_compare_qwen3.py

phase 10: add Qwen3-8B benchmark + performance fix

2026-05-22 10:25:33 +08:00

bench_compare.py

phase 8: add benchmark framework + baseline results

2026-05-21 23:29:41 +08:00

bench_gpt_oss.sh

server: support GptOss in TP engine + benchmark script

2026-05-30 15:39:44 +08:00

bench_server.py

tools: add correctness + performance test scripts for Qwen3-8B

2026-05-23 14:13:49 +08:00

bench_vs_hf.py

fix: comprehensive review + 14 bug fixes + Phase 12/14 overhaul

2026-05-22 17:53:28 +08:00

compare_logits.py

fix: comprehensive review + 14 bug fixes + Phase 12/14 overhaul

2026-05-22 17:53:28 +08:00

convert-to-gguf.sh

tools: add llama.cpp comparison baseline + standard benchmark suite

2026-05-28 11:18:52 +08:00

e2e_validate.py

fix: comprehensive review + 14 bug fixes + Phase 12/14 overhaul

2026-05-22 17:53:28 +08:00

eval_gsm8k_batch.sh

quantization: add FP8 E4M3 W8A16 for gpt-oss MoE expert weights

2026-06-07 19:33:07 +08:00

pp_diag.sh

bench: PP harness (xserv --pp vs llama.cpp -sm layer)

2026-05-29 18:45:59 +08:00

pp_final.sh

bench: PP harness (xserv --pp vs llama.cpp -sm layer)

2026-05-29 18:45:59 +08:00

pp_llama_47.sh

bench: PP harness (xserv --pp vs llama.cpp -sm layer)

2026-05-29 18:45:59 +08:00

pp_quality_full.sh

bench: PP harness (xserv --pp vs llama.cpp -sm layer)

2026-05-29 18:45:59 +08:00

pp_verify.sh

bench: PP harness (xserv --pp vs llama.cpp -sm layer)

2026-05-29 18:45:59 +08:00

quantize_fp8.py

quantization: add FP8 E4M3 W8A16 for gpt-oss MoE expert weights

2026-06-07 19:33:07 +08:00

run_gpt_oss_bench.sh

server: support GptOss in TP engine + benchmark script

2026-05-30 15:39:44 +08:00

setup-llama-cpp.sh

tools: add llama.cpp comparison baseline + standard benchmark suite

2026-05-28 11:18:52 +08:00

sync-and-build.sh

tools: add llama.cpp comparison baseline + standard benchmark suite

2026-05-28 11:18:52 +08:00

test_concurrent.py

fix: comprehensive review + 14 bug fixes + Phase 12/14 overhaul

2026-05-22 17:53:28 +08:00

test_correctness.py

tools: add correctness + performance test scripts for Qwen3-8B

2026-05-23 14:13:49 +08:00