Merge pull request #200 from danforbes/dfo/chore/refactors

Refactors from llm-chain integration

binaries/llm-cli/src/cli_args.rs

@@ -3,8 +3,8 @@ use std::{fmt::Debug, path::PathBuf};
 use clap::{Parser, Subcommand, ValueEnum};
 use color_eyre::eyre::{Result, WrapErr};
 use llm::{
-    ElementType, InferenceParameters, InferenceSessionConfig, LoadProgress, Model,
-    ModelKVMemoryType, ModelParameters, TokenBias,
+    ElementType, InferenceParameters, InferenceSessionConfig, InvalidTokenBias, LoadProgress,
+    Model, ModelKVMemoryType, ModelParameters, TokenBias,
 };
 use rand::SeedableRng;
 
@@ -276,7 +276,7 @@ impl Generate {
         }
     }
 }
-fn parse_bias(s: &str) -> Result<TokenBias, String> {
+fn parse_bias(s: &str) -> Result<TokenBias, InvalidTokenBias> {
     s.parse()
 }
 

crates/llm-base/src/lib.rs

@@ -32,7 +32,7 @@ pub use memmap2::Mmap;
 pub use model::{Hyperparameters, KnownModel, Model, ModelParameters, OutputRequest};
 pub use quantize::{quantize, QuantizeError, QuantizeProgress};
 pub use util::TokenUtf8Buffer;
-pub use vocabulary::{TokenBias, TokenId, Vocabulary};
+pub use vocabulary::{InvalidTokenBias, TokenBias, TokenId, Vocabulary};
 
 #[derive(Clone, Debug, PartialEq)]
 /// The parameters for text generation.

crates/llm-base/src/vocabulary.rs

@@ -1,4 +1,4 @@
-use std::{collections::HashMap, str::FromStr};
+use std::{collections::HashMap, error::Error, fmt::Display, str::FromStr};
 
 use crate::InferenceError;
 
@@ -139,7 +139,7 @@ impl TokenBias {
 }
 
 impl FromStr for TokenBias {
-    type Err = String;
+    type Err = InvalidTokenBias;
 
     /// A comma separated list of token biases. The list should be in the format
     /// "TID=BIAS,TID=BIAS" where TID is an integer token ID and BIAS is a
@@ -165,11 +165,30 @@ impl FromStr for TokenBias {
                     .map_err(|e: std::num::ParseFloatError| e.to_string())?;
                 Result::<_, String>::Ok((tid, bias))
             })
-            .collect::<Result<_, _>>()?;
+            .collect::<Result<_, _>>()
+            .map_err(InvalidTokenBias)?;
         Ok(TokenBias::new(x))
     }
 }
 
+/// An error was encountered when parsing a token bias string, which should be
+/// in the format "TID=BIAS,TID=BIAS" where TID is an integer token ID and BIAS
+/// is a floating point number.
+#[derive(Debug)]
+pub struct InvalidTokenBias(String);
+
+impl Display for InvalidTokenBias {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(
+            f,
+            "should be in the format <int>=<float>,<int>=<float>: {:?}",
+            self.0
+        )
+    }
+}
+
+impl Error for InvalidTokenBias {}
+
 impl std::fmt::Display for TokenBias {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
         write!(f, "{:?}", self.0)

crates/llm/Cargo.toml

@@ -15,6 +15,8 @@ llm-gptj = { path = "../models/gptj", optional = true, version = "0.1.1" }
 llm-bloom = { path = "../models/bloom", optional = true, version = "0.1.1" }
 llm-neox = { path = "../models/neox", optional = true, version = "0.1.1" }
 
+serde = { workspace = true }
+
 [dev-dependencies]
 rand = { workspace = true }
 

crates/llm/src/lib.rs

@@ -70,10 +70,11 @@ use std::{
 pub use llm_base::{
     ggml::format as ggml_format, load, load_progress_callback_stdout, quantize, ElementType,
     FileType, InferenceError, InferenceParameters, InferenceRequest, InferenceSession,
-    InferenceSessionConfig, InferenceSnapshot, KnownModel, LoadError, LoadProgress, Loader, Model,
-    ModelKVMemoryType, ModelParameters, OutputRequest, QuantizeError, QuantizeProgress,
-    SnapshotError, TokenBias, TokenId, TokenUtf8Buffer, Vocabulary,
+    InferenceSessionConfig, InferenceSnapshot, InvalidTokenBias, KnownModel, LoadError,
+    LoadProgress, Loader, Model, ModelKVMemoryType, ModelParameters, OutputRequest, QuantizeError,
+    QuantizeProgress, SnapshotError, TokenBias, TokenId, TokenUtf8Buffer, Vocabulary,
 };
+use serde::Serialize;
 
 /// All available models.
 pub mod models {
@@ -89,7 +90,7 @@ pub mod models {
     pub use llm_neox::{self as neox, NeoX};
 }
 
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Serialize)]
 /// All available model architectures.
 pub enum ModelArchitecture {
     #[cfg(feature = "bloom")]