llama.cpp
implement adaptive-p sampler
#17927
Merged

Commits
  • initial commit for branch
    ddh0 committed 102 days ago
  • simplify constants
    ddh0 committed 102 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 101 days ago
  • add params to `struct common_params_sampling`, add reference to PR
    ddh0 committed 101 days ago
  • explicitly clamp `min_target` and `max_target` to `[0.0, 1.0]`
    ddh0 committed 101 days ago
  • add args, rename `queue_size` -> `window_size`
    ddh0 committed 101 days ago
  • improved comments
    ddh0 committed 101 days ago
  • minor
    ddh0 committed 101 days ago
  • remove old unused code from algorithm
    ddh0 committed 101 days ago
  • minor
    ddh0 committed 101 days ago
  • add power law case to `common_sampler_init`, add sampler name mappings
    ddh0 committed 101 days ago
  • clarify behaviour when `window_size = 0`
    ddh0 committed 101 days ago
  • add missing enums
    ddh0 committed 101 days ago
  • remove `target_range` param, make `target == 1` no-op, cleanup code
    ddh0 committed 101 days ago
  • oops, straggler
    ddh0 committed 101 days ago
  • add missing parameters in `server-task.cpp`
    ddh0 committed 100 days ago
  • copy from author
    ddh0 committed 100 days ago
  • remove old debug log, style nit
    ddh0 committed 99 days ago
  • fix compiler warning, add commented-out logging per token
    ddh0 committed 99 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 99 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 99 days ago
  • re-write + change parameters + simplify
    ddh0 committed 99 days ago
  • oops forgot args.cpp
    ddh0 committed 99 days ago
  • fix leftover `window_size`
    ddh0 committed 99 days ago
  • add missing values to `common_params_sampling::print()`
    ddh0 committed 99 days ago
  • with logging
    ddh0 committed 98 days ago
  • does this fix it?
    ddh0 committed 98 days ago
  • no, but does this?
    ddh0 committed 98 days ago
  • update default decay
    ddh0 committed 98 days ago
  • Merge branch 'master' into power-law-sampler
    ddh0 committed 98 days ago
  • optimize
    ddh0 committed 98 days ago
  • fix bad merge
    ddh0 committed 98 days ago
  • silence `missing initializer for member`
    ddh0 committed 98 days ago
  • update default decay to 0.9
    ddh0 committed 98 days ago
  • fix logging
    ddh0 committed 98 days ago
  • format (double)
    ddh0 committed 98 days ago
  • add power law to the new `samplers` vector
    ddh0 committed 98 days ago
  • log sampler init values
    ddh0 committed 98 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 97 days ago
  • improve logging messages in llama_sampler_power_law
    ddh0 committed 97 days ago
  • remove extraneous logging
    ddh0 committed 97 days ago
  • simplify target computation
    ddh0 committed 97 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 97 days ago
  • remove debug logging, explicitly clamp params at init
    ddh0 committed 97 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 96 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 96 days ago
  • add `use_power_law` flag + logic, minor cleanup
    ddh0 committed 95 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 95 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 94 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 93 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 92 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 91 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 90 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 89 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 87 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 86 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 86 days ago
  • update `power-law` -> `adaptive-p`
    ddh0 committed 85 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 84 days ago
  • fix cold start EMA
    ddh0 committed 84 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 83 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 82 days ago
  • update `SHARPNESS` constant to `10.0f`
    ddh0 committed 82 days ago
  • minor style fixes
    ddh0 committed 82 days ago
  • minor style fixes cont.
    ddh0 committed 82 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 81 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 79 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 77 days ago
  • update `llama_sampler_adaptive_p_i` for backend sampling (ref: #17004)
    ddh0 committed 77 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 76 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 76 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 75 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 74 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 74 days ago
  • separate into `apply` + `accept` functions
    ddh0 committed 74 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 73 days ago
  • `pending_token_idx`: switch from `llama_token` to `int32`
    ddh0 committed 73 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 72 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 70 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 69 days ago
  • don't transform logits <= -1e9f
    ddh0 committed 69 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 68 days ago
  • fix masking in backend top-p, min-p
    ddh0 committed 68 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 68 days ago
  • address review comments
    ddh0 committed 67 days ago
  • typo in comments `RND` -> `RNG`
    ddh0 committed 67 days ago
  • add docs
    ddh0 committed 67 days ago
  • add recommended values in completion docs
    ddh0 committed 67 days ago
  • address PR feedback
    ddh0 committed 67 days ago
  • remove trailing whitespace (for CI `editorconfig`)
    ddh0 committed 67 days ago
  • Merge branch 'ggml-org:master' into power-law-sampler
    ddh0 committed 66 days ago
  • add to adaptive-p to `common_sampler_types_from_chars`
    ddh0 committed 66 days ago
Loading