vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

llama : save and restore kv cache for single seq id (#6341)

Browse source 
* llama : save and restore kv cache for single seq id

* remove trailing whitespace

* respond error in case there's no space in the kv cache

* add kv seq save restore to test case

* add --slot-save-path arg to enable save restore and restrict save location

* Returning 0 for some cases, instead of asserting.

* cleanup error cases

* rename sequence state functions

* rename state get set functions

* add previous function names back in with DEPRECATED notice

* update doc

* adjust endpoints to preferred style

* fix restoring zero cell count

* handle seq rm return value

* unused param

* keep in the size check

* fix return types

* add server test case for slot save restore

* cleanup

* add cake

* cleanup style

* add special

* removing a whole sequence never fails

* move sequence state file functionality from server to llama to match session api and add version tags

* catch exceptions on save as well

* error log messages

* check types for stricter restore

* update server doc

* readme : update API changes date

* strict filename validation

* move include, reject bom as well

* also reject empty filename

* reject whitespace and trailing dot

---------

Co-authored-by: Martin Evans <martindevans@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Jan Boon 2024-04-08 20:43:30 +08:00 • committed by GitHub
parent 87fb5b4234
commit beea6e1b16
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
11 changed files with 1086 additions and 31 deletions
Download patch file Download diff file Expand all files Collapse all files

58
examples/server/tests/features/slotsave.feature Normal file
View file

@ -0,0 +1,58 @@
@llama.cpp
@slotsave
Feature: llama.cpp server slot management
Background: Server startup
Given a server listening on localhost:8080
And a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models
And prompt caching is enabled
And 2 slots
And . as slot save path
And 2048 KV cache size
And 42 as server seed
And 24 max tokens to predict
Then the server is starting
Then the server is healthy
Scenario: Save and Restore Slot
# First prompt in slot 1 should be fully processed
Given a user prompt "What is the capital of France?"
And using slot id 1
And a completion request with no api error
Then 24 tokens are predicted matching (Lily|cake)
And 22 prompt tokens are processed
When the slot 1 is saved with filename "slot1.bin"
Then the server responds with status code 200
# Since we have cache, this should only process the last tokens
Given a user prompt "What is the capital of Germany?"
And a completion request with no api error
Then 24 tokens are predicted matching (Thank|special)
And 7 prompt tokens are processed
# Loading the original cache into slot 0,
# we should only be processing 1 prompt token and get the same output
When the slot 0 is restored with filename "slot1.bin"
Then the server responds with status code 200
Given a user prompt "What is the capital of France?"
And using slot id 0
And a completion request with no api error
Then 24 tokens are predicted matching (Lily|cake)
And 1 prompt tokens are processed
# For verification that slot 1 was not corrupted during slot 0 load, same thing
Given a user prompt "What is the capital of Germany?"
And using slot id 1
And a completion request with no api error
Then 24 tokens are predicted matching (Thank|special)
And 1 prompt tokens are processed
Scenario: Erase Slot
Given a user prompt "What is the capital of France?"
And using slot id 1
And a completion request with no api error
Then 24 tokens are predicted matching (Lily|cake)
And 22 prompt tokens are processed
When the slot 1 is erased
Then the server responds with status code 200
Given a user prompt "What is the capital of France?"
And a completion request with no api error
Then 24 tokens are predicted matching (Lily|cake)
And 22 prompt tokens are processed