From ec2be7bf57078f9500622d7c4a904d7b1bf992b6 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Mon, 28 Oct 2024 08:56:12 +0200 Subject: [PATCH] llama : remove Tail-Free sampling ggml-ci --- common/arg.cpp | 9 +- common/common.cpp | 1 - common/common.h | 4 +- common/sampling.cpp | 13 +-- examples/main/README.md | 8 -- examples/server/README.md | 11 +-- examples/server/public/index-new.html | 3 - examples/server/public/index.html | 2 - examples/server/server.cpp | 2 - examples/server/themes/buttons-top/index.html | 2 - examples/server/themes/wild/index.html | 2 - examples/server/utils.hpp | 2 +- include/llama.h | 3 - scripts/run-with-preset.py | 2 +- src/llama-sampling.cpp | 97 +------------------ tests/test-sampling.cpp | 26 +---- 16 files changed, 15 insertions(+), 172 deletions(-) diff --git a/common/arg.cpp b/common/arg.cpp index e1e933934f0ef..7c5c5e5cd5b88 100644 --- a/common/arg.cpp +++ b/common/arg.cpp @@ -943,13 +943,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.sparams.min_p = std::stof(value); } ).set_sparam()); - add_opt(common_arg( - {"--tfs"}, "N", - string_format("tail free sampling, parameter z (default: %.1f, 1.0 = disabled)", (double)params.sparams.tfs_z), - [](common_params & params, const std::string & value) { - params.sparams.tfs_z = std::stof(value); - } - ).set_sparam()); add_opt(common_arg( {"--xtc-probability"}, "N", string_format("xtc probability (default: %.1f, 0.0 = disabled)", (double)params.sparams.xtc_probability), @@ -1074,7 +1067,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex ).set_sparam()); add_opt(common_arg( {"--mirostat"}, "N", - string_format("use Mirostat sampling.\nTop K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.\n" + string_format("use Mirostat sampling.\nTop K, Nucleus and Locally Typical samplers are ignored if used.\n" "(default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)", params.sparams.mirostat), [](common_params & params, int value) { params.sparams.mirostat = value; diff --git a/common/common.cpp b/common/common.cpp index ff8cc4076e95d..7656843b116dd 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -2090,7 +2090,6 @@ void yaml_dump_non_result_info(FILE * stream, const common_params & params, cons const std::vector tensor_split_vector(params.tensor_split, params.tensor_split + llama_max_devices()); yaml_dump_vector_float(stream, "tensor_split", tensor_split_vector); - fprintf(stream, "tfs: %f # default: 1.0\n", sparams.tfs_z); fprintf(stream, "threads: %d # default: %u\n", params.cpuparams.n_threads, std::thread::hardware_concurrency()); fprintf(stream, "top_k: %d # default: 40\n", sparams.top_k); fprintf(stream, "top_p: %f # default: 0.95\n", sparams.top_p); diff --git a/common/common.h b/common/common.h index 18b2121ed89b0..cd5a8e051d33a 100644 --- a/common/common.h +++ b/common/common.h @@ -88,7 +88,7 @@ enum common_sampler_type { COMMON_SAMPLER_TYPE_TOP_K = 2, COMMON_SAMPLER_TYPE_TOP_P = 3, COMMON_SAMPLER_TYPE_MIN_P = 4, - COMMON_SAMPLER_TYPE_TFS_Z = 5, + //COMMON_SAMPLER_TYPE_TFS_Z = 5, COMMON_SAMPLER_TYPE_TYPICAL_P = 6, COMMON_SAMPLER_TYPE_TEMPERATURE = 7, COMMON_SAMPLER_TYPE_XTC = 8, @@ -113,7 +113,6 @@ struct common_sampler_params { float min_p = 0.05f; // 0.0 = disabled float xtc_probability = 0.00f; // 0.0 = disabled float xtc_threshold = 0.10f; // > 0.5 disables XTC - float tfs_z = 1.00f; // 1.0 = disabled float typ_p = 1.00f; // typical_p, 1.0 = disabled float temp = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities float dynatemp_range = 0.00f; // 0.0 = disabled @@ -139,7 +138,6 @@ struct common_sampler_params { std::vector samplers = { COMMON_SAMPLER_TYPE_DRY, COMMON_SAMPLER_TYPE_TOP_K, - COMMON_SAMPLER_TYPE_TFS_Z, COMMON_SAMPLER_TYPE_TYPICAL_P, COMMON_SAMPLER_TYPE_TOP_P, COMMON_SAMPLER_TYPE_MIN_P, diff --git a/common/sampling.cpp b/common/sampling.cpp index 48a9df8ba5b88..7922fde47d369 100644 --- a/common/sampling.cpp +++ b/common/sampling.cpp @@ -131,11 +131,11 @@ std::string common_sampler_params::print() const { snprintf(result, sizeof(result), "\trepeat_last_n = %d, repeat_penalty = %.3f, frequency_penalty = %.3f, presence_penalty = %.3f\n" "\tdry_multiplier = %.3f, dry_base = %.3f, dry_allowed_length = %d, dry_penalty_last_n = %d\n" - "\ttop_k = %d, tfs_z = %.3f, top_p = %.3f, min_p = %.3f, xtc_probability = %.3f, xtc_threshold = %.3f, typical_p = %.3f, temp = %.3f\n" + "\ttop_k = %d, top_p = %.3f, min_p = %.3f, xtc_probability = %.3f, xtc_threshold = %.3f, typical_p = %.3f, temp = %.3f\n" "\tmirostat = %d, mirostat_lr = %.3f, mirostat_ent = %.3f", penalty_last_n, penalty_repeat, penalty_freq, penalty_present, dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n, - top_k, tfs_z, top_p, min_p, xtc_probability, xtc_threshold, typ_p, temp, + top_k, top_p, min_p, xtc_probability, xtc_threshold, typ_p, temp, mirostat, mirostat_eta, mirostat_tau); return std::string(result); @@ -199,9 +199,6 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co case COMMON_SAMPLER_TYPE_XTC: llama_sampler_chain_add(result->chain, llama_sampler_init_xtc (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed)); break; - case COMMON_SAMPLER_TYPE_TFS_Z: - llama_sampler_chain_add(result->chain, llama_sampler_init_tail_free(params.tfs_z, params.min_keep)); - break; case COMMON_SAMPLER_TYPE_TYPICAL_P: llama_sampler_chain_add(result->chain, llama_sampler_init_typical (params.typ_p, params.min_keep)); break; @@ -373,7 +370,6 @@ char common_sampler_type_to_chr(enum common_sampler_type cnstr) { switch (cnstr) { case COMMON_SAMPLER_TYPE_DRY: return 'd'; case COMMON_SAMPLER_TYPE_TOP_K: return 'k'; - case COMMON_SAMPLER_TYPE_TFS_Z: return 'f'; case COMMON_SAMPLER_TYPE_TYPICAL_P: return 'y'; case COMMON_SAMPLER_TYPE_TOP_P: return 'p'; case COMMON_SAMPLER_TYPE_MIN_P: return 'm'; @@ -388,7 +384,6 @@ std::string common_sampler_type_to_str(enum common_sampler_type cnstr) { switch (cnstr) { case COMMON_SAMPLER_TYPE_DRY: return "dry"; case COMMON_SAMPLER_TYPE_TOP_K: return "top_k"; - case COMMON_SAMPLER_TYPE_TFS_Z: return "tfs_z"; case COMMON_SAMPLER_TYPE_TYPICAL_P: return "typ_p"; case COMMON_SAMPLER_TYPE_TOP_P: return "top_p"; case COMMON_SAMPLER_TYPE_MIN_P: return "min_p"; @@ -406,7 +401,6 @@ std::vector common_sampler_types_from_names(const std::vect { "top_p", COMMON_SAMPLER_TYPE_TOP_P }, { "typ_p", COMMON_SAMPLER_TYPE_TYPICAL_P }, { "min_p", COMMON_SAMPLER_TYPE_MIN_P }, - { "tfs_z", COMMON_SAMPLER_TYPE_TFS_Z }, { "temperature", COMMON_SAMPLER_TYPE_TEMPERATURE }, { "xtc", COMMON_SAMPLER_TYPE_XTC }, { "infill", COMMON_SAMPLER_TYPE_INFILL }, @@ -423,8 +417,6 @@ std::vector common_sampler_types_from_names(const std::vect { "typ-p", COMMON_SAMPLER_TYPE_TYPICAL_P }, { "typ", COMMON_SAMPLER_TYPE_TYPICAL_P }, { "min-p", COMMON_SAMPLER_TYPE_MIN_P }, - { "tfs-z", COMMON_SAMPLER_TYPE_TFS_Z }, - { "tfs", COMMON_SAMPLER_TYPE_TFS_Z }, { "temp", COMMON_SAMPLER_TYPE_TEMPERATURE }, }; @@ -452,7 +444,6 @@ std::vector common_sampler_types_from_chars(const std::stri std::unordered_map sampler_name_map = { { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_DRY), COMMON_SAMPLER_TYPE_DRY }, { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_K), COMMON_SAMPLER_TYPE_TOP_K }, - { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TFS_Z), COMMON_SAMPLER_TYPE_TFS_Z }, { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TYPICAL_P), COMMON_SAMPLER_TYPE_TYPICAL_P }, { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_P), COMMON_SAMPLER_TYPE_TOP_P }, { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_MIN_P), COMMON_SAMPLER_TYPE_MIN_P }, diff --git a/examples/main/README.md b/examples/main/README.md index c7c8231717980..5357ac2e2fa8f 100644 --- a/examples/main/README.md +++ b/examples/main/README.md @@ -235,14 +235,6 @@ The Min-P sampling method was designed as an alternative to Top-P, and aims to e Example usage: `--min-p 0.05` -### Tail-Free Sampling (TFS) - -- `--tfs N`: Enable tail free sampling with parameter z (default: 1.0, 1.0 = disabled). - -Tail-free sampling (TFS) is a text generation technique that aims to reduce the impact of less likely tokens, which may be less relevant, less coherent, or nonsensical, on the output. Similar to Top-P it tries to determine the bulk of the most likely tokens dynamically. But TFS filters out logits based on the second derivative of their probabilities. Adding tokens is stopped after the sum of the second derivatives reaches the parameter z. In short: TFS looks at how quickly the probabilities of the tokens decrease and cuts off the tail of unlikely tokens using the parameter z. Typical values for z are in the range of 0.9 to 0.95. A value of 1.0 would include all tokens and thus disables the effect of TFS. - -Example usage: `--tfs 0.95` - ### Locally Typical Sampling - `--typical N`: Enable locally typical sampling with parameter p (default: 1.0, 1.0 = disabled). diff --git a/examples/server/README.md b/examples/server/README.md index bc737237eb018..1629e456b6836 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -99,7 +99,7 @@ The project is under active development, and we are [looking for feedback and co | Argument | Explanation | | -------- | ----------- | -| `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'
(default: top_k;tfs_z;typ_p;top_p;min_p;temperature) | +| `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'
(default: top_k;typ_p;top_p;min_p;temperature) | | `-s, --seed SEED` | RNG seed (default: -1, use random seed for -1) | | `--sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: kfypmt) | | `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) | @@ -108,7 +108,6 @@ The project is under active development, and we are [looking for feedback and co | `--top-k N` | top-k sampling (default: 40, 0 = disabled) | | `--top-p N` | top-p sampling (default: 0.9, 1.0 = disabled) | | `--min-p N` | min-p sampling (default: 0.1, 0.0 = disabled) | -| `--tfs N` | tail free sampling, parameter z (default: 1.0, 1.0 = disabled) | | `--typical N` | locally typical sampling, parameter p (default: 1.0, 1.0 = disabled) | | `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) | | `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled) | @@ -121,7 +120,7 @@ The project is under active development, and we are [looking for feedback and co | `--dry-sequence-breaker STRING` | add sequence breaker for DRY sampling, clearing out default breakers (`['\n', ':', '"', '*']`) in the process; use `"none"` to not use any sequence breakers | `--dynatemp-range N` | dynamic temperature range (default: 0.0, 0.0 = disabled) | | `--dynatemp-exp N` | dynamic temperature exponent (default: 1.0) | -| `--mirostat N` | use Mirostat sampling.
Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.
(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) | +| `--mirostat N` | use Mirostat sampling.
Top K, Nucleus and Locally Typical samplers are ignored if used.
(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) | | `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.1) | | `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.0) | | `-l, --logit-bias TOKEN_ID(+/-)BIAS` | modifies the likelihood of token appearing in the completion,
i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',
or `--logit-bias 15043-1` to decrease likelihood of token ' Hello' | @@ -360,8 +359,6 @@ node index.js `stop`: Specify a JSON array of stopping strings. These words will not be included in the completion, so make sure to add them to the prompt for the next iteration. Default: `[]` - `tfs_z`: Enable tail free sampling with parameter z. Default: `1.0`, which is disabled. - `typical_p`: Enable locally typical sampling with parameter p. Default: `1.0`, which is disabled. `repeat_penalty`: Control the repetition of token sequences in the generated text. Default: `1.1` @@ -412,7 +409,7 @@ node index.js `cache_prompt`: Re-use KV cache from a previous request if possible. This way the common prefix does not have to be re-processed, only the suffix that differs between the requests. Because (depending on the backend) the logits are **not** guaranteed to be bit-for-bit identical for different batch sizes (prompt processing vs. token generation) enabling this option can cause nondeterministic results. Default: `false` - `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. Default: `["top_k", "tfs_z", "typical_p", "top_p", "min_p", "temperature"]` - these are all the available values. + `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. Default: `["top_k", "typical_p", "top_p", "min_p", "temperature"]` - these are all the available values. **Response format** @@ -738,7 +735,6 @@ Example: "repeat_penalty": 1.100000023841858, "samplers": [ "top_k", - "tfs_z", "typical_p", "top_p", "min_p", @@ -752,7 +748,6 @@ Example: "stream": false, "task_id": 0, "temperature": 0.0, - "tfs_z": 1.0, "top_k": 40, "top_p": 0.949999988079071, "typical_p": 1.0 diff --git a/examples/server/public/index-new.html b/examples/server/public/index-new.html index cb3995abee0ef..8bfa380e57388 100644 --- a/examples/server/public/index-new.html +++ b/examples/server/public/index-new.html @@ -49,7 +49,6 @@ min_p: 0.05, // 0 = disabled; recommended for non-english: ~ 0.4 xtc_probability: 0.0, // 0 = disabled; xtc_threshold: 0.1, // > 0.5 disables XTC; - tfs_z: 1.0, // 1.0 = disabled typical_p: 1.0, // 1.0 = disabled presence_penalty: 0.0, // 0.0 = disabled frequency_penalty: 0.0, // 0.0 = disabled @@ -847,7 +846,6 @@ ${FloatField({ label: "DRY Base", title: "Set the DRY repetition penalty base value. Default is 1.75", max: 3.0, min: 1.0, name: "dry_base", step: 0.01, value: params.value.dry_base })} ${IntField({ label: "DRY Allowed Length", title: "Tokens that extend repetition beyond this receive exponentially increasing penalty. Default is 2", max: 10, min: 1, step: 1, name: "dry_allowed_length", value: params.value.dry_allowed_length })} ${IntField({ label: "DRY Penalty Last N", title: "How many tokens to scan for repetitions. Default is -1, where 0 is disabled and -1 is context size", max: 2048, min: -1, step: 16, name: "dry_penalty_last_n", value: params.value.dry_penalty_last_n })} - ${FloatField({ label: "TFS-Z", title: "Activates tail-free sampling, a method used to limit the prediction of tokens that are too frequent. The parameter z controls the strength of this limitation. A value of 1.0 means that this function is deactivated.", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z })} ${IntField({ label: "Min Keep", title: "If greater than 0, samplers are forced to return N possible tokens at minimum. Default is 0", max: 10, min: 0, name: "min_keep", value: params.value.min_keep })} @@ -1147,7 +1145,6 @@

llama.cpp

xtc_probability: { snapValue: 0.0, snapRangeMultiplier: 4 }, xtc_threshold: { snapValue: 0.5, snapRangeMultiplier: 4 }, top_p: { snapValue: 1.0, snapRangeMultiplier: 4 }, - tfs_z: { snapValue: 1.0, snapRangeMultiplier: 4 }, typical_p: { snapValue: 1.0, snapRangeMultiplier: 4 }, repeat_penalty: { snapValue: 1.0, snapRangeMultiplier: 4 }, presence_penalty: { snapValue: 0.0, snapRangeMultiplier: 4 }, diff --git a/examples/server/public/index.html b/examples/server/public/index.html index 7f9b02bfbf83b..a95f5c6df8775 100644 --- a/examples/server/public/index.html +++ b/examples/server/public/index.html @@ -313,7 +313,6 @@ min_p: 0.05, // 0 = disabled xtc_probability: 0.0, // 0 = disabled; xtc_threshold: 0.1, // > 0.5 disables XTC; - tfs_z: 1.0, // 1.0 = disabled typical_p: 1.0, // 1.0 = disabled presence_penalty: 0.0, // 0.0 = disabled frequency_penalty: 0.0, // 0.0 = disabled @@ -1015,7 +1014,6 @@
More options
- ${FloatField({ label: "TFS-Z", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z })} ${FloatField({ label: "Typical P", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p })} ${FloatField({ label: "Presence penalty", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty })} ${FloatField({ label: "Frequency penalty", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty })} diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 077c7ad1adffb..7953b50659049 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -809,7 +809,6 @@ struct server_context { slot.sparams.min_p = json_value(data, "min_p", default_sparams.min_p); slot.sparams.xtc_probability = json_value(data, "xtc_probability", default_sparams.xtc_probability); slot.sparams.xtc_threshold = json_value(data, "xtc_threshold", default_sparams.xtc_threshold); - slot.sparams.tfs_z = json_value(data, "tfs_z", default_sparams.tfs_z); slot.sparams.typ_p = json_value(data, "typical_p", default_sparams.typ_p); slot.sparams.temp = json_value(data, "temperature", default_sparams.temp); slot.sparams.dynatemp_range = json_value(data, "dynatemp_range", default_sparams.dynatemp_range); @@ -1149,7 +1148,6 @@ struct server_context { {"min_p", slot.sparams.min_p}, {"xtc_probability", slot.sparams.xtc_probability}, {"xtc_threshold", slot.sparams.xtc_threshold}, - {"tfs_z", slot.sparams.tfs_z}, {"typical_p", slot.sparams.typ_p}, {"repeat_last_n", slot.sparams.penalty_last_n}, {"repeat_penalty", slot.sparams.penalty_repeat}, diff --git a/examples/server/themes/buttons-top/index.html b/examples/server/themes/buttons-top/index.html index 8334bcde5049c..2797c37c96456 100644 --- a/examples/server/themes/buttons-top/index.html +++ b/examples/server/themes/buttons-top/index.html @@ -226,7 +226,6 @@ top_k: 40, // <= 0 to use vocab size top_p: 0.95, // 1.0 = disabled min_p: 0.05, // 0 = disabled - tfs_z: 1.0, // 1.0 = disabled typical_p: 1.0, // 1.0 = disabled presence_penalty: 0.0, // 0.0 = disabled frequency_penalty: 0.0, // 0.0 = disabled @@ -788,7 +787,6 @@
More options
- ${FloatField({ label: "TFS-Z", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z })} ${FloatField({ label: "Typical P", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p })} ${FloatField({ label: "Presence penalty", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty })} ${FloatField({ label: "Frequency penalty", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty })} diff --git a/examples/server/themes/wild/index.html b/examples/server/themes/wild/index.html index 8361c577494d7..dbe23c4024155 100644 --- a/examples/server/themes/wild/index.html +++ b/examples/server/themes/wild/index.html @@ -229,7 +229,6 @@ top_k: 40, // <= 0 to use vocab size top_p: 0.95, // 1.0 = disabled min_p: 0.05, // 0 = disabled - tfs_z: 1.0, // 1.0 = disabled typical_p: 1.0, // 1.0 = disabled presence_penalty: 0.0, // 0.0 = disabled frequency_penalty: 0.0, // 0.0 = disabled @@ -791,7 +790,6 @@
More options
- ${FloatField({ label: "TFS-Z", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z })} ${FloatField({ label: "Typical P", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p })} ${FloatField({ label: "Presence penalty", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty })} ${FloatField({ label: "Frequency penalty", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty })} diff --git a/examples/server/utils.hpp b/examples/server/utils.hpp index 562635555e0ab..58f5a5684ac11 100644 --- a/examples/server/utils.hpp +++ b/examples/server/utils.hpp @@ -607,7 +607,7 @@ static json oaicompat_completion_params_parse( } // Copy remaining properties to llama_params - // This allows user to use llama.cpp-specific params like "mirostat", "tfs_z",... via OAI endpoint. + // This allows user to use llama.cpp-specific params like "mirostat", ... via OAI endpoint. // See "launch_slot_with_task()" for a complete list of params supported by llama.cpp for (const auto & item : body.items()) { // Exception: if "n_predict" is present, we overwrite the value specified earlier by "max_tokens" diff --git a/include/llama.h b/include/llama.h index b2d1e7d5ae16b..4076d34a7ae5a 100644 --- a/include/llama.h +++ b/include/llama.h @@ -1087,9 +1087,6 @@ extern "C" { /// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841 LLAMA_API struct llama_sampler * llama_sampler_init_min_p (float p, size_t min_keep); - /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/. - LLAMA_API struct llama_sampler * llama_sampler_init_tail_free (float z, size_t min_keep); - /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666. LLAMA_API struct llama_sampler * llama_sampler_init_typical (float p, size_t min_keep); diff --git a/scripts/run-with-preset.py b/scripts/run-with-preset.py index 47cacb4321589..8f0bf8ca8aa55 100755 --- a/scripts/run-with-preset.py +++ b/scripts/run-with-preset.py @@ -20,7 +20,7 @@ "np-penalize-nl", "numa", "ppl-output-type", "ppl-stride", "presence-penalty", "prompt", "prompt-cache", "prompt-cache-all", "prompt-cache-ro", "repeat-last-n", "repeat-penalty", "reverse-prompt", "rope-freq-base", "rope-freq-scale", "rope-scale", "seed", - "simple-io", "tensor-split", "threads", "temp", "tfs", "top-k", "top-p", "typical", + "simple-io", "tensor-split", "threads", "temp", "top-k", "top-p", "typical", "verbose-prompt" ] diff --git a/src/llama-sampling.cpp b/src/llama-sampling.cpp index 25536eb6c5a06..c2cfe0a77ad84 100644 --- a/src/llama-sampling.cpp +++ b/src/llama-sampling.cpp @@ -113,7 +113,7 @@ static void llama_sampler_softmax_impl(llama_token_data_array * cur_p) { } static void llama_sampler_top_k_impl(llama_token_data_array * cur_p, int32_t k) { - // TODO: move bucket sort to separate function so that top_p/tail_free/typical/softmax first is equally fast + // TODO: move bucket sort to separate function so that top_p/typical/softmax first is equally fast // if (k >= (int32_t)cur_p->size) { // return; // } @@ -733,101 +733,6 @@ struct llama_sampler * llama_sampler_init_min_p(float p, size_t min_keep) { }; } -// tail-free - -struct llama_sampler_tail_free { - const float z; - const size_t min_keep; -}; - -static const char * llama_sampler_tail_free_name(const struct llama_sampler * /*smpl*/) { - return "tail-free"; -} - -static void llama_sampler_tail_free_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) { - const auto * ctx = (llama_sampler_tail_free *) smpl->ctx; - - if (ctx->z >= 1.0f || cur_p->size <= 2) { - return; - } - - llama_sampler_softmax_impl(cur_p); - - // Compute the first and second derivatives - std::vector first_derivatives(cur_p->size - 1); - std::vector second_derivatives(cur_p->size - 2); - - for (size_t i = 0; i < first_derivatives.size(); ++i) { - first_derivatives[i] = cur_p->data[i].p - cur_p->data[i + 1].p; - } - for (size_t i = 0; i < second_derivatives.size(); ++i) { - second_derivatives[i] = first_derivatives[i] - first_derivatives[i + 1]; - } - - // Calculate absolute value of second derivatives - for (size_t i = 0; i < second_derivatives.size(); ++i) { - second_derivatives[i] = std::abs(second_derivatives[i]); - } - - // Normalize the second derivatives - { - const float second_derivatives_sum = std::accumulate(second_derivatives.begin(), second_derivatives.end(), 0.0f); - - if (second_derivatives_sum > 1e-6f) { - for (float & value : second_derivatives) { - value /= second_derivatives_sum; - } - } else { - for (float & value : second_derivatives) { - value = 1.0f / second_derivatives.size(); - } - } - } - - float cum_sum = 0.0f; - size_t last_idx = cur_p->size; - for (size_t i = 0; i < second_derivatives.size(); ++i) { - cum_sum += second_derivatives[i]; - - // Check if the running sum is greater than z or if we have kept at least min_keep tokens - if (cum_sum > ctx->z && i >= ctx->min_keep) { - last_idx = i; - break; - } - } - - // Resize the output vector to keep only the tokens above the tail location - cur_p->size = last_idx; -} - -static struct llama_sampler * llama_sampler_tail_free_clone(const struct llama_sampler * smpl) { - const auto * ctx = (const llama_sampler_tail_free *) smpl->ctx; - return llama_sampler_init_tail_free(ctx->z, ctx->min_keep); -} - -static void llama_sampler_tail_free_free(struct llama_sampler * smpl) { - delete (llama_sampler_tail_free *) smpl->ctx; -} - -static struct llama_sampler_i llama_sampler_tail_free_i = { - /* .name = */ llama_sampler_tail_free_name, - /* .accept = */ nullptr, - /* .apply = */ llama_sampler_tail_free_apply, - /* .reset = */ nullptr, - /* .clone = */ llama_sampler_tail_free_clone, - /* .free = */ llama_sampler_tail_free_free, -}; - -struct llama_sampler * llama_sampler_init_tail_free(float z, size_t min_keep) { - return new llama_sampler { - /* .iface = */ &llama_sampler_tail_free_i, - /* .ctx = */ new llama_sampler_tail_free { - /* .z = */ z, - /*. min_keep = */ min_keep, - }, - }; -} - // typical struct llama_sampler_typical { diff --git a/tests/test-sampling.cpp b/tests/test-sampling.cpp index eb39661c3698f..be370044d6f58 100644 --- a/tests/test-sampling.cpp +++ b/tests/test-sampling.cpp @@ -105,16 +105,6 @@ static void test_top_p(const std::vector & probs, const std::vector & probs, const std::vector & probs_expected, float z) { - sampler_tester tester(probs, probs_expected); - - DUMP(&tester.cur_p); - tester.apply(llama_sampler_init_tail_free(z, 1)); - DUMP(&tester.cur_p); - - tester.check(); -} - static void test_min_p(const std::vector & probs, const std::vector & probs_expected, float p) { sampler_tester tester(probs, probs_expected); @@ -202,7 +192,6 @@ static void test_sampler_queue(const size_t n_vocab, const std::string & sampler for (auto s : samplers_sequence) { switch (s){ case 'k': tester.apply(llama_sampler_init_top_k(top_k)); break; - case 'f': GGML_ABORT("tail_free test not implemented"); case 'y': GGML_ABORT("typical test not implemented"); case 'p': tester.apply(llama_sampler_init_top_p(top_p, 1)); break; case 'm': tester.apply(llama_sampler_init_min_p(min_p, 1)); break; @@ -299,12 +288,11 @@ static void test_perf() { data.emplace_back(llama_token_data{i, logit, 0.0f}); } - BENCH(llama_sampler_init_top_k (40), data, 32); - BENCH(llama_sampler_init_top_p (0.8f, 1), data, 32); - BENCH(llama_sampler_init_min_p (0.2f, 1), data, 32); - BENCH(llama_sampler_init_tail_free(0.5f, 1), data, 32); - BENCH(llama_sampler_init_typical (0.5f, 1), data, 32); - BENCH(llama_sampler_init_xtc (1.0f, 0.1f, 1, 1), data, 32); + BENCH(llama_sampler_init_top_k (40), data, 32); + BENCH(llama_sampler_init_top_p (0.8f, 1), data, 32); + BENCH(llama_sampler_init_min_p (0.2f, 1), data, 32); + BENCH(llama_sampler_init_typical(0.5f, 1), data, 32); + BENCH(llama_sampler_init_xtc (1.0f, 0.1f, 1, 1), data, 32); } int main(void) { @@ -343,10 +331,6 @@ int main(void) { printf("XTC should not:\n"); test_xtc({0.4f, 0.3f, 0.2f, 0.1f}, {0.4f, 0.3f, 0.2f, 0.1f}, 0.99f, 0.39f); - test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f}, 0.25f); - test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f}, 0.75f); - test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f}, 0.99f); - test_typical({0.97f, 0.01f, 0.01f, 0.01f}, {0.97f}, 0.5f); test_typical({0.4f, 0.2f, 0.2f, 0.2f}, {0.2f, 0.2f, 0.2f}, 0.5f);