1
2# llama-gguf-hash
3
4CLI to hash GGUF files to detect difference on a per model and per tensor level.
5
6**Command line options:**
7
8- `--help`: display help message
9- `--xxh64`: use xhash 64bit hash mode (default)
10- `--sha1`: use sha1
11- `--uuid`: use uuid
12- `--sha256`: use sha256
13- `--all`: use all hash
14- `--no-layer`: exclude per layer hash
15- `--uuid`: generate UUIDv5 ID
16- `-c`, `--check <manifest>`: verify against a manifest
17
18## About
19
20While most POSIX systems already have hash checking programs like sha256sum, it
21is designed to check entire files. This is not ideal for our purpose if we want
22to check for consistency of the tensor data even if the metadata content of the
23gguf KV store has been updated.
24
25This program is designed to hash a gguf tensor payload on a 'per tensor layer'
26in addition to a 'entire tensor model' hash. The intent is that the entire
27tensor layer can be checked first but if there is any detected inconsistencies,
28then the per tensor hash can be used to narrow down the specific tensor layer
29that has inconsistencies.
30
31For Maintainers:
32- Detection of tensor inconsistency during development and automated tests
33 - This is served by xxh64 which is fast
34 - This is also served by having per tensor layer to assist in narrowing down
35 the location of the faulty tensor layer
36 - This is also served by sha1 which is much slower but more widely supported
37
38For Model Creators:
39- Optional consistent UUID generation based on model tensor content
40 - This is served by UUIDv5 which is useful for databases keys
41 - llama.cpp UUIDv5 Namespace: `ef001206-dadc-5f6d-a15f-3359e577d4e5`
42 - Made via UUIDv5 URL namespace of `en.wikipedia.org/wiki/Llama.cpp`
43
44For Model Users:
45- Assurance of tensor layer integrity even if metadata was updated
46 - This is served by sha256 which is still considered very secure as of 2024
47
48### Design Note
49
50- The default behavior of this program if no arguments is provided is to hash
51 using xxhash's xxh32 mode because it is very fast and is primarily targeted
52 towards maintainers who may want to use this in automated tests.
53- xxhash support xxh32 and xxh128 for 32bit hash and 128bit hash respectively
54 however we picked 64bit xxhash as most computers are 64bit as of 2024 and thus
55 would have a better affinity to calculating hash that is 64bit in size.
56
57## Compile Example
58
59```bash
60cmake -B build -DCMAKE_BUILD_TYPE=Debug -DLLAMA_FATAL_WARNINGS=ON
61make -C build clean
62make -C build llama-gguf-hash VERBOSE=1
63./build/bin/llama-gguf-hash test.gguf
64./build/bin/llama-gguf-hash --xxh64 test.gguf
65./build/bin/llama-gguf-hash --sha1 test.gguf
66./build/bin/llama-gguf-hash --uuid test.gguf
67./build/bin/llama-gguf-hash --sha256 test.gguf
68```
69
70## Generation and Verification Example
71
72To generate we may use this command
73
74```bash
75./llama-gguf-hash --all test.gguf > test.gguf.manifest
76```
77
78Which would generate a manifest that looks like below, which contains multiple hash type and per tensor layer hashes as well
79(This excludes UUID as that is an ID not a hash)
80
81```bash
82xxh64 f66e9cd66a4396a0 test.gguf:tensor_0
83sha1 59f79ecefd8125a996fdf419239051a7e99e5f20 test.gguf:tensor_0
84sha256 c0510d38fa060c46265e0160a85c7243096b01dd31c2f355bdbb5516b20de1bd test.gguf:tensor_0
85xxh64 7d3a1f9ac04d0537 test.gguf:tensor_1
86sha1 4765f592eacf096df4628ba59476af94d767080a test.gguf:tensor_1
87sha256 8514cbcc73692a2c56bd7a33a022edd5ff819614bd23b19915d7224387f397a7 test.gguf:tensor_1
88xxh64 a0af5d700049693b test.gguf:tensor_2
89sha1 25cbfbad4513cc348e2c95ebdee69d6ff2fd8753 test.gguf:tensor_2
90sha256 947e6b36e20f2cc95e1d2ce1c1669d813d574657ac6b5ac5196158d454d35180 test.gguf:tensor_2
91xxh64 e83fddf559d7b6a6 test.gguf:tensor_3
92sha1 a9cba73e2d90f2ee3dae2548caa42bef3fe6a96c test.gguf:tensor_3
93sha256 423b044e016d8ac73c39f23f60bf01bedef5ecb03c0230accd824c91fe86f1a1 test.gguf:tensor_3
94xxh64 1257733306b7992d test.gguf:tensor_4
95sha1 d7bc61db93bb685ce9d598da89717c66729b7543 test.gguf:tensor_4
96sha256 79737cb3912d4201384cf7f16a1a37ff7823f23ea796cb205b6ca361ab9e3ebf test.gguf:tensor_4
97xxh64 d238d16ba4711e58 test.gguf:tensor_5
98sha1 0706566c198fe1072f37e0a5135b4b5f23654c52 test.gguf:tensor_5
99sha256 60949be8298eced0ecdde64487643d018407bd261691e061d9e9c3dbc9fd358b test.gguf:tensor_5
100xxh64 3fbc3b65ab8c7f39 test.gguf:tensor_6
101sha1 73922a0727226a409049f6fc3172a52219ca6f00 test.gguf:tensor_6
102sha256 574f4c46ff384a3b9a225eb955d2a871847a2e8b3fa59387a8252832e92ef7b0 test.gguf:tensor_6
103xxh64 c22021c29854f093 test.gguf:tensor_7
104sha1 efc39cece6a951188fc41e354c73bbfe6813d447 test.gguf:tensor_7
105sha256 4c0410cd3c500f078ae5b21e8dc9eb79e29112713b2ab58a882f82a3868d4d75 test.gguf:tensor_7
106xxh64 936df61f5d64261f test.gguf:tensor_8
107sha1 c2490296d789a4f34398a337fed8377d943d9f06 test.gguf:tensor_8
108sha256 c4401313feeba0261275c3b25bd2d8fe40ce04e0f440c2980ed0e9674c30ff01 test.gguf:tensor_8
109xxh64 93fd20c64421c081 test.gguf:tensor_9
110sha1 7047ce1e78437a6884337a3751c7ee0421918a65 test.gguf:tensor_9
111sha256 23d57cf0d7a6e90b0b3616b41300e0cd354781e812add854a5f95aa55f2bc514 test.gguf:tensor_9
112xxh64 5a54d3aad816f302 test.gguf
113sha1 d15be52c4ff213e823cb6dd13af7ee2f978e7042 test.gguf
114sha256 7dd641b32f59b60dbd4b5420c4b0f6321ccf48f58f6ae201a3dbc4a58a27c6e4 test.gguf
115```
116
117We can then use the normal check command which will by default check for the highest security strength hash and verify against that:
118
119```bash
120$ ./llama-gguf-hash --check test.gguf.manifest test.gguf
121manifest test.gguf.manifest sha256 sha1 xxh64
122sha256 c0510d38fa060c46265e0160a85c7243096b01dd31c2f355bdbb5516b20de1bd test.gguf:tensor_0 - Ok
123sha256 8514cbcc73692a2c56bd7a33a022edd5ff819614bd23b19915d7224387f397a7 test.gguf:tensor_1 - Ok
124sha256 947e6b36e20f2cc95e1d2ce1c1669d813d574657ac6b5ac5196158d454d35180 test.gguf:tensor_2 - Ok
125sha256 423b044e016d8ac73c39f23f60bf01bedef5ecb03c0230accd824c91fe86f1a1 test.gguf:tensor_3 - Ok
126sha256 79737cb3912d4201384cf7f16a1a37ff7823f23ea796cb205b6ca361ab9e3ebf test.gguf:tensor_4 - Ok
127sha256 60949be8298eced0ecdde64487643d018407bd261691e061d9e9c3dbc9fd358b test.gguf:tensor_5 - Ok
128sha256 574f4c46ff384a3b9a225eb955d2a871847a2e8b3fa59387a8252832e92ef7b0 test.gguf:tensor_6 - Ok
129sha256 4c0410cd3c500f078ae5b21e8dc9eb79e29112713b2ab58a882f82a3868d4d75 test.gguf:tensor_7 - Ok
130sha256 c4401313feeba0261275c3b25bd2d8fe40ce04e0f440c2980ed0e9674c30ff01 test.gguf:tensor_8 - Ok
131sha256 23d57cf0d7a6e90b0b3616b41300e0cd354781e812add854a5f95aa55f2bc514 test.gguf:tensor_9 - Ok
132sha256 7dd641b32f59b60dbd4b5420c4b0f6321ccf48f58f6ae201a3dbc4a58a27c6e4 test.gguf - Ok
133
134Verification results for test.gguf.manifest - Success
135```
136
137Or we may explicitly ask for a faster hash like:
138
139```bash
140$ ./llama-gguf-hash --check test.gguf.manifest --xxh64 test.gguf
141manifest test.gguf.manifest sha256 sha1 xxh64
142xxh64 f66e9cd66a4396a0 test.gguf:tensor_0 - Ok
143xxh64 7d3a1f9ac04d0537 test.gguf:tensor_1 - Ok
144xxh64 a0af5d700049693b test.gguf:tensor_2 - Ok
145xxh64 e83fddf559d7b6a6 test.gguf:tensor_3 - Ok
146xxh64 1257733306b7992d test.gguf:tensor_4 - Ok
147xxh64 d238d16ba4711e58 test.gguf:tensor_5 - Ok
148xxh64 3fbc3b65ab8c7f39 test.gguf:tensor_6 - Ok
149xxh64 c22021c29854f093 test.gguf:tensor_7 - Ok
150xxh64 936df61f5d64261f test.gguf:tensor_8 - Ok
151xxh64 93fd20c64421c081 test.gguf:tensor_9 - Ok
152xxh64 5a54d3aad816f302 test.gguf - Ok
153
154Verification results for test.gguf.manifest - Success
155```
156
157Or maybe we want to just check that all the hash is valid:
158
159```bash
160$./llama-gguf-hash --check test.gguf.manifest --all test.gguf.manifest
161manifest test.gguf.manifest sha256 sha1 xxh64
162xxh64 f66e9cd66a4396a0 test.gguf:tensor_0 - Ok
163sha1 59f79ecefd8125a996fdf419239051a7e99e5f20 test.gguf:tensor_0 - Ok
164sha256 c0510d38fa060c46265e0160a85c7243096b01dd31c2f355bdbb5516b20de1bd test.gguf:tensor_0 - Ok
165xxh64 7d3a1f9ac04d0537 test.gguf:tensor_1 - Ok
166sha1 4765f592eacf096df4628ba59476af94d767080a test.gguf:tensor_1 - Ok
167sha256 8514cbcc73692a2c56bd7a33a022edd5ff819614bd23b19915d7224387f397a7 test.gguf:tensor_1 - Ok
168xxh64 a0af5d700049693b test.gguf:tensor_2 - Ok
169sha1 25cbfbad4513cc348e2c95ebdee69d6ff2fd8753 test.gguf:tensor_2 - Ok
170sha256 947e6b36e20f2cc95e1d2ce1c1669d813d574657ac6b5ac5196158d454d35180 test.gguf:tensor_2 - Ok
171xxh64 e83fddf559d7b6a6 test.gguf:tensor_3 - Ok
172sha1 a9cba73e2d90f2ee3dae2548caa42bef3fe6a96c test.gguf:tensor_3 - Ok
173sha256 423b044e016d8ac73c39f23f60bf01bedef5ecb03c0230accd824c91fe86f1a1 test.gguf:tensor_3 - Ok
174xxh64 1257733306b7992d test.gguf:tensor_4 - Ok
175sha1 d7bc61db93bb685ce9d598da89717c66729b7543 test.gguf:tensor_4 - Ok
176sha256 79737cb3912d4201384cf7f16a1a37ff7823f23ea796cb205b6ca361ab9e3ebf test.gguf:tensor_4 - Ok
177xxh64 d238d16ba4711e58 test.gguf:tensor_5 - Ok
178sha1 0706566c198fe1072f37e0a5135b4b5f23654c52 test.gguf:tensor_5 - Ok
179sha256 60949be8298eced0ecdde64487643d018407bd261691e061d9e9c3dbc9fd358b test.gguf:tensor_5 - Ok
180xxh64 3fbc3b65ab8c7f39 test.gguf:tensor_6 - Ok
181sha1 73922a0727226a409049f6fc3172a52219ca6f00 test.gguf:tensor_6 - Ok
182sha256 574f4c46ff384a3b9a225eb955d2a871847a2e8b3fa59387a8252832e92ef7b0 test.gguf:tensor_6 - Ok
183xxh64 c22021c29854f093 test.gguf:tensor_7 - Ok
184sha1 efc39cece6a951188fc41e354c73bbfe6813d447 test.gguf:tensor_7 - Ok
185sha256 4c0410cd3c500f078ae5b21e8dc9eb79e29112713b2ab58a882f82a3868d4d75 test.gguf:tensor_7 - Ok
186xxh64 936df61f5d64261f test.gguf:tensor_8 - Ok
187sha1 c2490296d789a4f34398a337fed8377d943d9f06 test.gguf:tensor_8 - Ok
188sha256 c4401313feeba0261275c3b25bd2d8fe40ce04e0f440c2980ed0e9674c30ff01 test.gguf:tensor_8 - Ok
189xxh64 93fd20c64421c081 test.gguf:tensor_9 - Ok
190sha1 7047ce1e78437a6884337a3751c7ee0421918a65 test.gguf:tensor_9 - Ok
191sha256 23d57cf0d7a6e90b0b3616b41300e0cd354781e812add854a5f95aa55f2bc514 test.gguf:tensor_9 - Ok
192xxh64 5a54d3aad816f302 test.gguf - Ok
193sha1 d15be52c4ff213e823cb6dd13af7ee2f978e7042 test.gguf - Ok
194sha256 7dd641b32f59b60dbd4b5420c4b0f6321ccf48f58f6ae201a3dbc4a58a27c6e4 test.gguf - Ok
195
196Verification results for test.gguf.manifest - Success
197```
198
199
200## Crypto/Hash Libraries Used
201
202These micro c libraries dependencies was installed via the [clib c package manager](https://github.com/clibs)
203
204- https://github.com/Cyan4973/xxHash
205- https://github.com/clibs/sha1/
206- https://github.com/jb55/sha256.c