32ac3153e4
default speculative set to 8. added more adapter fields
2024-11-30 16:18:27 +08:00
e0c59486ee
default to 12 tokens drafted
2024-11-30 11:52:07 +08:00
b21d0fe3ac
customizable speculative size
2024-11-30 11:28:19 +08:00
f75bbb945f
speculative decoding initial impl completed (+6 squashed commit)
...
Squashed commit:
[0a6306ca0] draft wip dont use (will be squashed)
[a758a1c9c] wip dont use (will be squashed)
[e1994d3ce] wip dont use
[f59690d68] wip
[77228147d] wip on spec decoding. dont use yet
[2445bca54] wip adding speculative decoding (+1 squashed commits)
Squashed commits:
[50e341bb7] wip adding speculative decoding
2024-11-30 10:41:10 +08:00
b9e99c69e8
fixed build
2024-11-26 22:06:55 +08:00
62dde8cfb2
ollama sync completions mostly working. stupid api.
2024-11-23 23:31:37 +08:00
2c1a06a07d
wip ollama emulation, added detokenize endpoint
2024-11-23 22:48:03 +08:00
1dd37933e3
fixed grammar not resetting correctly
2024-11-23 09:55:12 +08:00
547ab2aebb
API: add /props route ( #1222 )
...
* API: add an /extra/chat_template route
A lot of manual tweaking is done when swapping between models. We can automate or make better assumptions about some of them by having more information, such as chat template. This PR adds an endpoint /extra/chat_template which returns the model chat template string as is in a 'chat_template' key. The front end can then use this to derive the proper templates or use it as is, or at least warn the user when they are trying to use e.g. a Mistral preset with a Llama 3.1 model.
* switch to pre-established /props endpoint for chat template
* bug-fix (upstream): one-off in string juggling
2024-11-21 10:58:32 +08:00
70aee82552
attempts a backflip, but does he stick the landing?
2024-11-16 17:05:45 +08:00
bfa118ee45
fix llava segfault
2024-11-14 14:16:39 +08:00
4b96c3bba8
try new batch api (not actually batching)
2024-11-14 13:47:26 +08:00
3813f6c517
added new flag nofastforward allowing users to disable fast forwarding
2024-11-13 10:59:01 +08:00
48e9372337
prevent outputting infinity to logprobs (+1 squashed commits)
...
Squashed commits:
[bcc5f8b92] prevent outputting infinity to logprobs
2024-11-13 00:09:53 +08:00
3c36bbdcd7
debug: display tokens that were dropped by XTC sampler when debugmode is enabled ( #1201 )
2024-11-06 23:09:28 +08:00
223c5f0844
clblast survived
2024-11-02 21:51:38 +08:00
bbebc76817
fix top picks bug, lower input anti abuse thresholds (+1 squashed commits)
...
Squashed commits:
[a81d9b21] fix top picks bug, lower input anti abuse thresholds
2024-11-01 16:42:13 +08:00
aa26a58085
added logprobs api and logprobs viewer
2024-11-01 00:22:15 +08:00
90f5cd0f67
wip logprobs data
2024-10-30 00:59:34 +08:00
94a5a27b85
Alone in the darkness
...
They're coming for you
I know they will try to catch me too
Alone in the darkness
They're calling for you
There's nowhere to run for cover
2024-10-24 22:29:20 +08:00
becd737e0f
slightly increase padding to handle longer gen amts
2024-10-23 22:58:41 +08:00
8bb220329c
Dynamic sizes for sequences ( #1157 )
...
* Dynamic sizes for sequences
* cleanup PR - move all dynamic fields to end of payload, ensure correct null handling to match existing behavior, add anti abuse limit of max 512 for dynamic fields
* adjust anti abuse limits
---------
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com >
2024-10-16 23:55:11 +08:00
7f76425450
lower topk prefilter token amount to 3k
2024-10-16 20:39:41 +08:00
cff72c5d26
remove unwanted print
2024-10-11 18:56:32 +08:00
e692a79aab
Merge branch 'upstream' into concedo_experimental
...
# Conflicts:
# .github/workflows/docker.yml
# CMakeLists.txt
# CONTRIBUTING.md
# docs/android.md
# docs/docker.md
# examples/embedding/embedding.cpp
# examples/imatrix/imatrix.cpp
# examples/infill/infill.cpp
# examples/llama-bench/llama-bench.cpp
# examples/main/README.md
# examples/parallel/parallel.cpp
# examples/perplexity/perplexity.cpp
# examples/quantize-stats/quantize-stats.cpp
# examples/save-load-state/save-load-state.cpp
# examples/server/README.md
# examples/simple/CMakeLists.txt
# examples/speculative/speculative.cpp
# flake.lock
# ggml/src/CMakeLists.txt
# ggml/src/ggml-blas.cpp
# pocs/vdot/q8dot.cpp
# pocs/vdot/vdot.cpp
# scripts/debug-test.sh
# scripts/sync-ggml.last
# src/llama.cpp
# tests/test-backend-ops.cpp
# tests/test-chat-template.cpp
# tests/test-quantize-fns.cpp
# tests/test-quantize-perf.cpp
# tests/test-tokenizer-0.cpp
# tests/test-tokenizer-1-bpe.cpp
# tests/test-tokenizer-1-spm.cpp
2024-10-11 11:59:59 +08:00
5c9650d68e
Fix access violation when using banned_phrases ( #1154 )
2024-10-10 21:46:39 +08:00
fe5479f286
unify antislop and token bans
2024-10-10 18:21:07 +08:00
9b614d46bd
antislop sampler working
2024-10-09 16:33:04 +08:00
36e9bac98f
wip anti slop sampler
2024-10-09 13:34:47 +08:00
f78f8d3d45
wip anti slop
2024-10-07 23:18:13 +08:00
65f3c68399
wip antislop
2024-10-07 20:19:22 +08:00
740c5e01cb
added token delay feature
2024-10-07 19:45:51 +08:00
3e8bb10e2d
wip on rewind function
2024-10-06 16:21:03 +08:00
c38d1ecc8d
update templates, fix rwkv
2024-09-22 01:32:12 +08:00
53bf0fb32d
removed openblas backend, merged into CPU (with llamafile for BLAS). GPU backend is now automatically selected when running from CLI unless noblas is specified.
2024-09-15 19:21:52 +08:00
e44ddf26ef
Merge branch 'upstream' into concedo_experimental
...
# Conflicts:
# .github/workflows/build.yml
# .github/workflows/server.yml
# CMakeLists.txt
# Makefile
# examples/embedding/embedding.cpp
# examples/imatrix/imatrix.cpp
# examples/llama-bench/llama-bench.cpp
# examples/llava/MobileVLM-README.md
# examples/parallel/parallel.cpp
# examples/perplexity/perplexity.cpp
# examples/quantize/CMakeLists.txt
# examples/server/README.md
# examples/speculative/speculative.cpp
# tests/test-backend-ops.cpp
2024-09-13 16:17:24 +08:00
7bdac9bc44
prevent shifting on rwkv
2024-09-11 20:22:45 +08:00
eee67281be
move kcpp params out
2024-09-10 16:30:12 +08:00
fc7fe2e7a0
allow rwkv6 to run although its broken
2024-09-09 20:50:58 +08:00
b63158005f
All samplers moved to kcpp side
2024-09-09 18:14:11 +08:00
12fd16bfd4
Merge commit 'df270ef74596da8f1178f08991f4c51f18c9ee82' into concedo_experimental
...
# Conflicts:
# Makefile
# common/CMakeLists.txt
# common/common.h
# common/sampling.cpp
# common/sampling.h
# examples/infill/infill.cpp
# examples/llama-bench/llama-bench.cpp
# examples/quantize-stats/quantize-stats.cpp
# examples/server/server.cpp
# include/llama.h
# src/llama-sampling.cpp
# src/llama-sampling.h
# src/llama.cpp
# tests/test-grammar-integration.cpp
# tests/test-grammar-parser.cpp
# tests/test-json-schema-to-grammar.cpp
# tests/test-llama-grammar.cpp
# tests/test-sampling.cpp
2024-09-09 17:10:08 +08:00
c78690737c
fix for DRY segfault on unicode character substring tokenization
2024-09-08 18:25:00 +08:00
d220495dd4
Merge branch 'upstream' into concedo_experimental
...
# Conflicts:
# .devops/full-cuda.Dockerfile
# .devops/llama-cli-cuda.Dockerfile
# .devops/llama-server-cuda.Dockerfile
# .devops/llama-server-intel.Dockerfile
# .devops/llama-server-rocm.Dockerfile
# .devops/llama-server-vulkan.Dockerfile
# .devops/llama-server.Dockerfile
# .github/workflows/docker.yml
# docs/docker.md
# examples/llama-bench/llama-bench.cpp
# flake.lock
# ggml/include/ggml.h
# ggml/src/CMakeLists.txt
# scripts/sync-ggml.last
# src/llama.cpp
# tests/test-backend-ops.cpp
# tests/test-grad0.cpp
# tests/test-rope.cpp
2024-08-30 10:37:39 +08:00
b78a637da5
try to optimize context shifting
2024-08-26 23:07:31 +08:00
cca3c4c78b
xtc fixes
2024-08-22 23:18:46 +08:00
fc2545dc83
fixed a typo
2024-08-22 00:25:56 +08:00
5bf527a6ae
added xtc sampler
2024-08-21 23:57:15 +08:00
1a7ecd55e6
timing for init step, clip for vulkan
2024-08-21 18:14:53 +08:00
cd69ab218e
fixed DRY
2024-08-21 17:01:28 +08:00
6a4becb731
dry is still buggy because token indexes are wrong
2024-08-21 00:59:26 +08:00
Concedo
kallewoof
Maya