From 6f493220fc58e3727ca0ff37645fe2995d8643f1 Mon Sep 17 00:00:00 2001 From: Nam Nguyen Date: Mon, 8 Apr 2024 15:01:09 -0700 Subject: [PATCH] Add FIM config for CodeGemma Also remove some spaces. --- src/configTemplates.ts | 30 ++++++++++++++++++++++++++---- 1 file changed, 26 insertions(+), 4 deletions(-) diff --git a/src/configTemplates.ts b/src/configTemplates.ts index c19e42b..7c79668 100644 --- a/src/configTemplates.ts +++ b/src/configTemplates.ts @@ -1,4 +1,4 @@ -const templateKeys = ["hf/bigcode/starcoder", "hf/codellama/CodeLlama-13b-hf", "hf/Phind/Phind-CodeLlama-34B-v2", "hf/WizardLM/WizardCoder-Python-34B-V1.0", "ollama/codellama:7b", "hf/deepseek-ai/deepseek-coder-6.7b-base", "Custom"] as const; +const templateKeys = ["hf/bigcode/starcoder", "hf/codellama/CodeLlama-13b-hf", "hf/Phind/Phind-CodeLlama-34B-v2", "hf/WizardLM/WizardCoder-Python-34B-V1.0", "ollama/codellama:7b", "hf/deepseek-ai/deepseek-coder-6.7b-base", "hf/google/codegemma-2b", "Custom"] as const; export type TemplateKey = typeof templateKeys[number]; export interface TokenizerPathConfig { @@ -49,6 +49,27 @@ const HfStarCoderConfig: Config = { repository: "bigcode/starcoder", } } +const HfCodeGemmaConfig: Config = { + modelId: "google/codegemma-2b", + backend: "huggingface", + url: null, + "fillInTheMiddle.enabled": true, + "fillInTheMiddle.prefix": "<|fim_prefix|>", + "fillInTheMiddle.suffix": "<|fim_suffix|>", + "fillInTheMiddle.middle": "<|fim_middle|>", + contextWindow: 8192, + tokensToClear: ["<|file_separator|>"], + tokenizer: { + repository: "google/codegemma-2b", + }, + requestBody: { + parameters: { + max_new_tokens: 128, + temperature: 0.1, + top_p: 0.95 + } + } +} const HfCodeLlama13BConfig: Config = { modelId: "codellama/CodeLlama-13b-hf", @@ -78,11 +99,11 @@ const HfDeepSeekConfig: Config = { url: null, "fillInTheMiddle.enabled": true, "fillInTheMiddle.prefix": "<｜fim▁begin｜>", - // DeepSeek names the suffix token fim_hole, + // DeepSeek names the suffix token fim_hole, // as it indicates the position to fill in "fillInTheMiddle.suffix": "<｜fim▁hole｜>", "fillInTheMiddle.middle": "<｜fim▁end｜>", - // DeepSeek should support 16k, + // DeepSeek should support 16k, // keeping at 8k because of resource constraints contextWindow: 8192, tokensToClear: ["<|EOT|>"], @@ -131,9 +152,10 @@ const OllamaCodeLlama7BConfig: Config = { export const templates: Partial> = { "hf/bigcode/starcoder": HfStarCoderConfig, + "hf/google/codegemma-2b": HfCodeGemmaConfig, "hf/codellama/CodeLlama-13b-hf": HfCodeLlama13BConfig, "hf/Phind/Phind-CodeLlama-34B-v2": HfPhindCodeLlama34Bv2Config, "hf/WizardLM/WizardCoder-Python-34B-V1.0": HfWizardCoderPython34Bv1Config, - "hf/deepseek-ai/deepseek-coder-6.7b-base": HfDeepSeekConfig, + "hf/deepseek-ai/deepseek-coder-6.7b-base": HfDeepSeekConfig, "ollama/codellama:7b": OllamaCodeLlama7BConfig, }