Hugging Face's logo

Datasets:
yil384
/
sift-archive

Modalities:
Text
Languages:
Chinese
English
Size:
10B<n<100B
Tags:
Mixture of Experts
inference
numa
cxl
tiered-memory
kv-cache
License:
Dataset card Data Studio Files
xet
 Community
Dataset Preview
Duplicate
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed
Error code:   DatasetGenerationError
Exception:    UnicodeDecodeError
Message:      'utf-8' codec can't decode bytes in position 67957-67958: invalid continuation byte
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1827, in _prepare_split_single
                  for key, table in generator:
                                    ^^^^^^^^^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 613, in wrapped
                  for item in generator(*args, **kwargs):
                              ~~~~~~~~~^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/text/text.py", line 98, in _generate_tables
                  batch = f.read(self.config.chunksize)
                File "/usr/local/lib/python3.14/site-packages/datasets/utils/file_utils.py", line 844, in read_with_retries
                  out = read(*args, **kwargs)
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 732, in track_read
                  out = f_read(*args, **kwargs)
                File "<frozen codecs>", line 325, in decode
              UnicodeDecodeError: 'utf-8' codec can't decode bytes in position 67957-67958: invalid continuation byte
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
                  parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
                                                                        ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      builder, max_dataset_size_bytes=max_dataset_size_bytes
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
                  builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
                  ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1694, in _prepare_split
                  for job_id, done, content in self._prepare_split_single(
                                               ~~~~~~~~~~~~~~~~~~~~~~~~~~^
                      gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  ):
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1880, in _prepare_split_single
                  raise DatasetGenerationError("An error occurred while generating the dataset") from e
              datasets.exceptions.DatasetGenerationError: An error occurred while generating the dataset

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

text
string
Total number of tokens: 7126
The Project Gutenberg eBook of War and Peace, by Leo Tolstoy
This eBook is for the use of anyone anywhere in the United States and
most other parts of the world at no cost and with almost no restrictions
whatsoever. You may copy it, give it away or re-use it under the terms
of the Project Gutenberg License included with this eBook or online at
www.gutenberg.org. If you are not located in the United States, you
will have to check the laws of the country where you are located before
using this eBook.
Title: War and Peace
Author: Leo Tolstoy
Translators: Louise and Aylmer Maude
Release Date: April, 2001 [eBook #2600]
[Most recently updated: June 14, 2022]
Language: English
Character set encoding: UTF-8
Produced by: An Anonymous Volunteer and David Widger
*** START OF THE PROJECT GUTENBERG EBOOK WAR AND PEACE ***
WAR AND PEACE
By Leo Tolstoy/Tolstoi
Contents
BOOK ONE: 1805
CHAPTER I
CHAPTER II
CHAPTER III
CHAPTER IV
CHAPTER V
CHAPTER VI
CHAPTER VII
CHAPTER VIII
CHAPTER IX
CHAPTER X
CHAPTER XI
CHAPTER XII
CHAPTER XIII
CHAPTER XIV
CHAPTER XV
CHAPTER XVI
CHAPTER XVII
CHAPTER XVIII
CHAPTER XIX
CHAPTER XX
CHAPTER XXI
CHAPTER XXII
CHAPTER XXIII
CHAPTER XXIV
CHAPTER XXV
CHAPTER XXVI
CHAPTER XXVII
CHAPTER XXVIII
BOOK TWO: 1805
End of preview.

Sift — 研究数据归档

Sift 是一个 CPU/DDR-primary + GPU-assisted 的分层内存 MoE + 长上下文推理系统研究项目 （用便宜的大容量 DDR/CXL 承载放不进 HBM 的大型稀疏 MoE + 长上下文；头条指标是 tokens-per-dollar / tokens-per-Joule）。

本仓是该项目自产实验数据的归档，用于把数据从本地磁盘卸下来。 这里没有模型权重 —— 模型是上游公开 GGUF，见 MODELS.manifest.json + restore_models.sh。

取数据 

hf download yil384/sift-archive --repo-type dataset fetch_archive.sh --local-dir .
bash fetch_archive.sh                      # 列出仓里有什么
bash fetch_archive.sh ssd2/traces/v2lite   # 取一个目录（默认回原始绝对路径）
bash fetch_archive.sh ssd2 nvme            # 全量 ~40 GB

取模型（不在本仓，直取上游） 

hf download yil384/sift-archive --repo-type dataset MODELS.manifest.json restore_models.sh --local-dir .
bash restore_models.sh              # 全部 5 个 GGUF，~100 GB -> /nvme2n1/yichen/sift/models
bash restore_models.sh v2lite-q4    # 只取一个

restore_models.sh 用 manifest 里 pin 死的 revision SHA 下载并逐文件 sha256 校验，保证与当初实验用的字节完全一致。 manifest 于 2026-09-22 逐文件验证过上游仍可取（5/5 PASS）。

内容

前缀 文件 大小 内容
ssd2/traces/inputs 39 2.3 MB M1 trace 输入语料：wikitext / pg19 / gsm8k / humaneval 全文 + 64 KiB chunk 切片
ssd2/traces/v2lite 15 5.36 GiB M1 专家路由 trace — DeepSeek-V2-Lite Q4_K_M
ssd2/traces/v2lite_d1 4 6.33 GiB M1 Δ=1 预取窗口实验 trace
ssd2/traces/qwen3 14 3.41 GiB M1 专家路由 trace — Qwen3-30B-A3B Q4_K_M（9.5× 偏斜、热度秩稳定 0.90–0.96 的来源）
ssd2/traces/mixtral 14 0.11 GiB M1 专家路由 trace — Mixtral-8x7B Q4_K_M（「L14 64%」主张未复现、实测 29–32% 的来源）
ssd2/traces/r1 4 0.06 GiB M1/M2 专家路由 trace — DeepSeek-R1 671B Q4_K_M ⚠ 不可重采（见下）
ssd2/traces/gateA_smoke.bin 1 0.40 GiB 门 A 冒烟 trace
nvme/traces 25 17.52 GiB M3 KQ trace 扩到 16K / 32K 上下文（mixtral/qwen3/v2lite × 4 workload × 2 深度）
nvme/qkdump_niah 54 4.27 GiB M3b NIAH 场景 Q/K 原始 dump（每次 attention 调用一个 81 MiB 文件）
nvme/qkdump_wiki 27 2.14 GiB M3b wikitext 场景 Q/K dump，泛化交叉验证用
nvme/qkdump 4 316 KB M3b Q/K dump 的 prompt / token / 日志
nvme/m3b_pred 12 41 MB M3b 预测器实验：niah_trace.bin + 分析脚本副本
nvme/tmp 4 6.9 MB M3b 中间件：v2lite_interim.predbasis（SVD 基）、NIAH 失败样例
合计 217 39.64 GiB

ARCHIVE_MANIFEST.json 里有逐文件的 sha256 + 原始绝对路径 + 产出脚本。

两处需要知道的坑

  1. ssd2/traces/r1 不可重采。 它的源模型 DeepSeek-R1 671B Q4_K_M（404 GB）已于 2026-08-31 被本地清理删除， 且清理脚本未记录上游 repo id。M2 里程碑的 671B 数字与 results/m2/r1.place 放置表（7.1× 偏斜）都源于这份 trace。
  2. Intel MLC v3.12 未收录。 /nvme2n1/yichen/sift/tools/mlc 的 redist.txt 声明二进制受 Intel EULA 约束， 故不转发到公开仓；可从 Intel 官网重取（far-memory 延迟标定用，见 scripts/far_calibrate.sh）。

输入语料来源

ssd2/traces/inputs 派生自公开数据集 WikiText、PG-19、GSM8K、HumanEval，由 scripts/make_trace_inputs.py 切片产生。 二进制 trace 记录的是模型内部的专家路由决策与 Q/K 统计，不含上述公开语料以外的内容。
Downloads last month
60
Total file size:
42.6 GB