[Feat] Add UCM store compression module and compression config parametersFeature store compress (#940)

## Purpose
This PR introduces a configurable compression module for the UCM store,
designed to reduce **TTFT (Time-To-First-Token)** under SSD hit. By
compressing stored BF16 tensors and supporting multi-threaded
decompression, aim to lower I/O overhead and speed up cache access,
improving end-to-end inference latency.

## Modifications 
- Added three new configuration parameters under ucm_connector_config to
control compression behavior:
- `compress_ratio`: Compression ratio (currently only 2.0x is supported)
  - `data_type`: Target data type (currently only BF16 is supported)
- `decompress_thread_num`: Number of threads used for parallel
decompression
- Implemented full compression/decompression pipeline based on KVfold
coding
- Integrated the compression module into the existing UCM store pipeline
- Added YAML configuration support for compression feature toggling and
tuning

## Test

- Model: Qwen2.5-14B-Instruct
- Hardware: Kunpeng 920 5250 + 4 × Ascend 910B4 (tensor parallel = 4)
- Test Config: 0% hit on DRAM cache, 50% / 80% / 100% hit on local SSD.

Layer_wise (use_layerwise: true) :

input tokens | output tokens | batch size | odirect | TTFT/ms, 50% hit,
w/o compression | TTFT/ms, 50% hit, with compression | TTFT reduction |
TTFT/ms, 80% hit, w/o compression | TTFT/ms, 80% hit, with compression |
TTFT reduction | TTFT/ms, 100% hit, w/o compression | TTFT/ms, 100% hit,
with compression | TTFT reduction
-- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | --
4000 | 1 | 1 | TRUE | 357.99 | 341.97 | 4.47% | 207.02 | 225.74 | -9.04%
| 208.37 | 193.22 | 7.27%
8000 | 1 | 1 | TRUE | 531.6 | 578.86 | -8.89% | 304.06 | 321.29 | -5.67%
| 350.01 | 331.22 | 5.37%
16000 | 1 | 1 | TRUE | 1236.44 | 1284.19 | -3.86% | 800.25 | 720.4 |
9.98% | 611.74 | 532.57 | 12.94%
32000 | 1 | 1 | TRUE | 3278.62 | 3152.37 | 3.85% | 2077.46 | 1807.59 |
12.99% | 1204.62 | 959.57 | 20.34%
4000 | 1 | 8 | TRUE | 1339.06 | 1396.03 | -4.25% | 868.46 | 839.52 |
3.33% | 1056.99 | 829.89 | 21.49%
8000 | 1 | 8 | TRUE | 2353.23 | 2373.6 | -0.87% | 1436.33 | 1356.39 |
5.57% | 2024.04 | 1421.53 | 29.77%
16000 | 1 | 8 | TRUE | 5311.28 | 5548.89 | -4.47% | 3433.49 | 2829.16 |
17.60% | 3808.22 | 2906.72 | 23.67%
32000 | 1 | 8 | TRUE | 13796.04 | 13660.42 | 0.98% | 8353.08 | 7851.27 |
6.01% | 6939.55 | 5274.13 | 24.00%
4000 | 1 | 16 | TRUE | 2280.48 | 2343.98 | -2.78% | 1428.27 | 1232.16 |
13.73% | 2154.65 | 1342.75 | 37.68%
8000 | 1 | 16 | TRUE | 4414.05 | 4436.29 | -0.50% | 2639.71 | 2388.07 |
9.53% | 4170.57 | 2646 | 36.56%
16000 | 1 | 16 | TRUE | 10144.97 | 9960.42 | 1.82% | 6167.56 | 5084.08 |
17.57% | 7170.8 | 4005.21 | 44.15%
32000 | 1 | 16 | TRUE | 25967.02 | 24714.92 | 4.82% | 15550.23 |
12950.47 | 16.72% | 12261.77 | 7504.62 | 38.80%

Block_wise (use_layerwise: false) :

input tokens | output tokens | batch size | odirect | TTFT/ms, 50% hit,
w/o compression | TTFT/ms, 50% hit, with compression | TTFT reduction |
TTFT/ms, 80% hit, w/o compression | TTFT/ms, 80% hit, with compression |
TTFT reduction | TTFT/ms, 100% hit, w/o compression | TTFT/ms, 100% hit,
with compression | TTFT reduction
-- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | --
4000 | 1 | 1 | TRUE | 423.87 | 406.85 | 4.02% | 307.49 | 253.9 | 17.43%
| 260.84 | 198.13 | 24.04%
8000 | 1 | 1 | TRUE | 688.38 | 665.83 | 3.28% | 507.52 | 419.18 | 17.41%
| 460.22 | 352.85 | 23.33%
16000 | 1 | 1 | TRUE | 1528.11 | 1479.79 | 3.16% | 1083.55 | 874.61 |
19.28% | 617.35 | 446.64 | 27.65%
32000 | 1 | 1 | TRUE | 3678.05 | 3426.42 | 6.84% | 2327.42 | 2039.48 |
12.37% | 1387.85 | 832.33 | 40.03%
4000 | 1 | 8 | TRUE | 1801.28 | 1673.41 | 7.10% | 1261.3 | 1025.12 |
18.73% | 1117.65 | 745.11 | 33.33%
8000 | 1 | 8 | TRUE | 3262.38 | 2952.3 | 9.50% | 2344.01 | 1880.82 |
19.76% | 2381.43 | 1314.63 | 44.80%
16000 | 1 | 8 | TRUE | 6982.66 | 6343.26 | 9.16% | 4897.66 | 3862.16 |
21.14% | 4157.12 | 2060.17 | 50.44%
32000 | 1 | 8 | TRUE | 16615.58 | 14960.37 | 9.96% | 10797.82 | 8443.65
| 21.80% | 7872.14 | 3896.58 | 50.50%
4000 | 1 | 16 | TRUE | 3181.24 | 2928.33 | 7.95% | 2228.06 | 1775.39 |
20.32% | 2402.1 | 1152.3 | 52.03%
8000 | 1 | 16 | TRUE | 6082.47 | 5437.54 | 10.60% | 4213.28 | 3197.17 |
24.12% | 4381.51 | 2386.19 | 45.54%
16000 | 1 | 16 | TRUE | 12832.2 | 11805.53 | 8.00% | 8489.46 | 6721.34 |
20.83% | 6231.51 | 3555.04 | 42.95%
32000 | 1 | 16 | TRUE | 29728.23 | 28228.75 | 5.04% | 18745.21 |
15293.23 | 18.42% | 13077.52 | 6831.96 | 47.76%

Co-authored-by: xwLearnsLLM <wangxuan154@huawei.com>

Author: fanzhust

.pre-commit-config.yaml

@@ -6,7 +6,7 @@ repos:
         exclude: \.(jsonl|txt)$
         args: [
           '--skip', 'ucm/csrc/**,ucm.egg-info/**,.github/**,ucm/sparse/gsa_on_device/csrc/**',
-          '-L', 'CANN,cann,NNAL,nnal,ASCEND,ascend,EnQue,CopyIn'
+          '-L', 'CANN,cann,NNAL,nnal,ASCEND,ascend,EnQue,CopyIn,Collet,DElt,re-use'
         ]
   - repo: https://github.com/psf/black
     rev: 24.4.2

ucm/store/CMakeLists.txt

@@ -5,10 +5,11 @@ target_link_libraries(storeintf INTERFACE storedetail infra_status)
 add_subdirectory(nfsstore)
 add_subdirectory(pcstore)
 add_subdirectory(ds3fs)
+add_subdirectory(posix)
+add_subdirectory(compress)
 add_subdirectory(cache)
 add_subdirectory(empty)
 add_subdirectory(fake)
 add_subdirectory(mooncakestore)
 add_subdirectory(pipeline)
-add_subdirectory(posix)
 add_subdirectory(test)

ucm/store/compress/CMakeLists.txt

@@ -0,0 +1,11 @@
+file(GLOB_RECURSE UCM_COMPRESSOR_STORE_CC_SOURCE_FILES "./cc/*.cc")
+
+add_library(compressor SHARED ${UCM_COMPRESSOR_STORE_CC_SOURCE_FILES})
+target_include_directories(compressor PUBLIC
+    ${CMAKE_CURRENT_SOURCE_DIR}/cc
+    ${CMAKE_CURRENT_SOURCE_DIR}/cc/compress_lib
+)
+target_link_libraries(compressor PUBLIC storeintf infra_logger)
+
+file(RELATIVE_PATH INSTALL_REL_PATH ${CMAKE_SOURCE_DIR} ${CMAKE_CURRENT_SOURCE_DIR})
+install(TARGETS compressor LIBRARY DESTINATION ${INSTALL_REL_PATH} COMPONENT ucm)