date timestamp[s]date 2026-08-22 00:00:00 2026-08-31 00:00:00 | engine stringclasses 6
values | model stringclasses 4
values | quant stringclasses 6
values | tp_size int64 0 2 | moe_backend stringclasses 3
values | attention_backend stringclasses 3
values | extra_flags stringclasses 4
values | ram_mts int64 2k 3.6k | category stringclasses 4
values | metric stringclasses 5
values | target stringclasses 9
values | value unknown | value_max float64 9.47 106 ⌀ | unit stringclasses 5
values | notes stringlengths 0 99 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
2026-08-22T00:00:00 | v0.1.2 | nvidia/Qwen3.6-35B-A3B-FP8 | fp8 | 1 | offload | hybrid_linear | 2,000 | serving | decode_tok_s | 93.9 | null | tok/s | warm; 399 tokens; single RTX 3090 | ||
2026-08-22T00:00:00 | v0.1.2 | nvidia/Qwen3.6-35B-A3B-FP8 | fp8 | 1 | offload | hybrid_linear | 2,000 | serving | ttft_s | 4.03 | null | s | warm | ||
2026-08-22T00:00:00 | v0.1.2 | nvidia/Qwen3.6-35B-A3B-FP8 | fp8 | 1 | offload | hybrid_linear | 2,000 | serving | decode_tok_s | 90.5 | null | tok/s | cold; includes FlashInfer JIT | ||
2026-08-22T00:00:00 | v0.1.2 | nvidia/Qwen3.6-35B-A3B-FP8 | fp8 | 1 | offload | hybrid_linear | 2,000 | serving | ttft_s | 79.4 | null | s | cold; JIT compile dominated | ||
2026-08-22T00:00:00 | v0.1.2 | nvidia/Qwen3.6-35B-A3B-NVFP4 | nvfp4 | 1 | hybrid | hybrid_linear | 2,000 | serving | decode_tok_s | 61 | null | tok/s | warm; 399 tokens | ||
2026-08-22T00:00:00 | v0.1.2 | nvidia/Qwen3.6-35B-A3B-NVFP4 | nvfp4 | 1 | hybrid | hybrid_linear | 2,000 | serving | ttft_s | 2.2800000000000002 | null | s | warm | ||
2026-08-22T00:00:00 | v0.1.2 | nvidia/Qwen3.6-35B-A3B-NVFP4 | nvfp4 | 1 | hybrid | hybrid_linear | 2,000 | serving | decode_tok_s | 56.4 | null | tok/s | cold | ||
2026-08-22T00:00:00 | v0.1.2 | nvidia/Qwen3.6-35B-A3B-FP8 | fp8 | 1 | offload | hybrid_linear | 3,600 | serving | decode_tok_s | 86 | 106 | tok/s | range across batches; flat vs 2000 MT/s — PCIe-bound, no fp8 CPU path | ||
2026-08-22T00:00:00 | v0.1.2 | nvidia/Qwen3.6-35B-A3B-NVFP4 | nvfp4 | 1 | hybrid | hybrid_linear | 3,600 | serving | decode_tok_s | 76 | 80.9 | tok/s | +29-31% vs 2000 MT/s — CPU executor benefits from RAM bandwidth | ||
2026-08-27T00:00:00 | PR#104 3c8b281 + f1ff8e2 | nvidia/Qwen3.6-35B-A3B-NVFP4 | nvfp4 | 2 | hybrid | triton | --tensor-parallel-size 2 --attention-backend triton --expert-load serial --disable-pynccl --memory-ratio 0.7 | 3,600 | serving | decode_tok_s | 2.2 | null | tok/s | pre-affinity-fix; both CPU executors pinned to same cores 0..16 | |
2026-08-27T00:00:00 | PR#104 3c8b281 + f1ff8e2 | nvidia/Qwen3.6-35B-A3B-NVFP4 | nvfp4 | 2 | hybrid | triton | --tensor-parallel-size 2 --attention-backend triton --expert-load serial --disable-pynccl --memory-ratio 0.7 --moe-cpu-threads 8 | 3,600 | serving | decode_tok_s | 8.62 | 9.47 | tok/s | workaround: cap threads per rank; 4.3x over broken | |
2026-08-27T00:00:00 | PR#104 3c8b281 + f1ff8e2 | nvidia/Qwen3.6-35B-A3B-NVFP4 | nvfp4 | 2 | offload | triton | --tensor-parallel-size 2 --attention-backend triton --expert-load serial --disable-pynccl --memory-ratio 0.7 | 3,600 | serving | decode_tok_s | 3.1 | null | tok/s | clean rerun; 219 tokens in 70.3s; earlier contaminated measurement (stale server on GPUs) discarded | |
2026-08-27T00:00:00 | PR#104 3c8b281 + f1ff8e2 | nvidia/Qwen3.6-35B-A3B-FP8 | fp8 | 2 | offload | triton | --tensor-parallel-size 2 --attention-backend triton --expert-load serial --disable-pynccl --memory-ratio 0.7 | 3,600 | load_test | load_status | "oom_at_load" | null | 62 GB RAM insufficient for ~35 GB pinned expert banks + runtime | ||
2026-08-26T00:00:00 | PR#104 3c8b281 | Qwen3.6-35B-A3B-BF16 | bf16 | 2 | offload | triton | --tensor-parallel-size 2 --attention-backend triton --expert-load serial --disable-pynccl --memory-ratio 0.7 | 3,600 | load_test | load_status | "oom_at_load" | null | 67 GB model vs 62 GB RAM; crashed at expert load (60/62 GB used) | ||
2026-08-28T00:00:00 | PR#104 + f1ff8e2 (lm_head fix) | nvidia/Qwen3.6-35B-A3B-FP8 | fp8 | 2 | hybrid | triton | --tensor-parallel-size 2 --attention-backend triton --expert-load serial --disable-pynccl --memory-ratio 0.7 | 3,600 | load_test | load_status | "pass" | null | lm_head vocab-sharding fix verified; passes previous crash point | ||
2026-08-28T00:00:00 | PR#104 + f1ff8e2 (lm_head fix) | nvidia/Qwen3.6-35B-A3B-NVFP4 | nvfp4 | 2 | hybrid | triton | --tensor-parallel-size 2 --attention-backend triton --expert-load serial --disable-pynccl --memory-ratio 0.7 | 3,600 | load_test | load_status | "pass" | null | lm_head vocab-sharding fix verified; passes previous crash point | ||
2026-08-31T00:00:00 | PR#104 + c23ec78 (core-partition fix) | nvidia/Qwen3.6-35B-A3B-NVFP4 | nvfp4 | 2 | hybrid | triton | --tensor-parallel-size 2 --attention-backend triton --expert-load serial --disable-pynccl --memory-ratio 0.7 | 3,600 | serving | decode_tok_s | 32.7 | null | tok/s | client wall; 219 tokens in 6.7s; disjoint pins rank0 0..14 / rank1 1..15 | |
2026-08-31T00:00:00 | PR#104 + c23ec78 (core-partition fix) | nvidia/Qwen3.6-35B-A3B-NVFP4 | nvfp4 | 2 | hybrid | triton | --tensor-parallel-size 2 --attention-backend triton --expert-load serial --disable-pynccl --memory-ratio 0.7 | 3,600 | serving | decode_tok_s | 59.26 | null | tok/s | engine-instantaneous per-batch gen throughput | |
2026-08-27T00:00:00 | PR#104 + f1ff8e2 | nvidia/Qwen3.6-35B-A3B-NVFP4 | nvfp4 | 1 | hybrid | triton | 3,600 | capacity | kv_usable_tokens | 262144 | null | tokens | baseline | ||
2026-08-27T00:00:00 | PR#104 + f1ff8e2 | nvidia/Qwen3.6-35B-A3B-NVFP4 | nvfp4 | 2 | hybrid | triton | --tensor-parallel-size 2 --attention-backend triton --expert-load serial --disable-pynccl --memory-ratio 0.7 | 3,600 | capacity | kv_usable_tokens | 524288 | null | tokens | 2.4x KV claim confirmed at TP=2 | |
2026-08-22T00:00:00 | v0.1.2 | 0 | ft bench bw | 2,000 | bandwidth | bandwidth_gbs | cpu_stream_read | 54.1 | null | GB/s | |||||
2026-08-22T00:00:00 | v0.1.2 | 0 | ft bench bw | 3,600 | bandwidth | bandwidth_gbs | cpu_stream_read | 80.8 | null | GB/s | |||||
2026-08-22T00:00:00 | v0.1.2 | 0 | ft bench bw | 2,000 | bandwidth | bandwidth_gbs | pcie_h2d | 8.6 | null | GB/s | |||||
2026-08-22T00:00:00 | v0.1.2 | 0 | ft bench bw | 3,600 | bandwidth | bandwidth_gbs | pcie_h2d | 8.5 | null | GB/s | |||||
2026-08-22T00:00:00 | v0.1.2 | 0 | ft bench bw | 2,000 | bandwidth | bandwidth_gbs | cpu_moe_nvfp4 | 28.1 | null | GB/s | |||||
2026-08-22T00:00:00 | v0.1.2 | 0 | ft bench bw | 3,600 | bandwidth | bandwidth_gbs | cpu_moe_nvfp4 | 46.7 | null | GB/s | |||||
2026-08-22T00:00:00 | v0.1.2 | 0 | ft bench bw | 2,000 | bandwidth | bandwidth_gbs | cpu_moe_bf16 | 41.5 | null | GB/s | |||||
2026-08-22T00:00:00 | v0.1.2 | 0 | ft bench bw | 3,600 | bandwidth | bandwidth_gbs | cpu_moe_bf16 | 71.3 | null | GB/s | |||||
2026-08-22T00:00:00 | v0.1.2 | 0 | ft bench bw | 2,000 | bandwidth | bandwidth_gbs | cpu_moe_ds_fp4 | 33.4 | null | GB/s | |||||
2026-08-22T00:00:00 | v0.1.2 | 0 | ft bench bw | 3,600 | bandwidth | bandwidth_gbs | cpu_moe_ds_fp4 | 49.2 | null | GB/s | |||||
2026-08-22T00:00:00 | v0.1.2 | 0 | ft bench bw | 2,000 | bandwidth | bandwidth_gbs | cpu_moe_mxfp4 | 21.2 | null | GB/s | |||||
2026-08-22T00:00:00 | v0.1.2 | 0 | ft bench bw | 3,600 | bandwidth | bandwidth_gbs | cpu_moe_mxfp4 | 35.2 | null | GB/s | |||||
2026-08-22T00:00:00 | v0.1.2 | bf16 | 0 | ft bench bw | 2,000 | bandwidth | bandwidth_gbs | cpu_moe_effective | 50.5 | null | GB/s | auto-picked backend: hybrid | |||
2026-08-22T00:00:00 | v0.1.2 | bf16 | 0 | ft bench bw | 2,000 | bandwidth | bandwidth_gbs | pcie_gather | 6.6 | null | GB/s | auto-picked backend: hybrid | |||
2026-08-22T00:00:00 | v0.1.2 | nvfp4 | 0 | ft bench bw | 2,000 | bandwidth | bandwidth_gbs | cpu_moe_effective | 48.6 | null | GB/s | auto-picked backend: hybrid | |||
2026-08-22T00:00:00 | v0.1.2 | nvfp4 | 0 | ft bench bw | 2,000 | bandwidth | bandwidth_gbs | pcie_gather | 6.9 | null | GB/s | auto-picked backend: hybrid | |||
2026-08-22T00:00:00 | v0.1.2 | fp8 | 0 | ft bench bw | 2,000 | bandwidth | bandwidth_gbs | pcie_gather | 8 | null | GB/s | auto-picked backend: offload | |||
2026-08-22T00:00:00 | v0.1.2 | mxfp4 | 0 | ft bench bw | 2,000 | bandwidth | bandwidth_gbs | cpu_moe_effective | 28.8 | null | GB/s | auto-picked backend: hybrid | |||
2026-08-22T00:00:00 | v0.1.2 | mxfp4 | 0 | ft bench bw | 2,000 | bandwidth | bandwidth_gbs | pcie_gather | 6 | null | GB/s | auto-picked backend: hybrid | |||
2026-08-22T00:00:00 | v0.1.2 | ds_fp4 | 0 | ft bench bw | 2,000 | bandwidth | bandwidth_gbs | cpu_moe_effective | 47.7 | null | GB/s | auto-picked backend: hybrid | |||
2026-08-22T00:00:00 | v0.1.2 | ds_fp4 | 0 | ft bench bw | 2,000 | bandwidth | bandwidth_gbs | pcie_gather | 6.4 | null | GB/s | auto-picked backend: hybrid |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.