geoffsee/predict-otron-9001
1
0
Fork 0
mirror of https://github.com/geoffsee/predict-otron-9001.git synced 2025-09-08 22:46:44 +00:00
Code Issues Packages Projects Releases Wiki Activity

- Change default server host to localhost for improved security.

Browse Source 
- Increase default maximum tokens in CLI configuration to 256.
- Refactor and reorganize CLI
This commit is contained in:
geoffsee
2025-08-27 21:47:24 -04:00
parent 766d41af78
commit 719beb3791
20 changed files with 1703 additions and 490 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
Cargo.lock generated
View File

@@ -2736,6 +2736,7 @@ dependencies = [
"symphonia",
"tokenizers",
"tokio",
"tokio-stream",
"tower",
"tower-http",
"tracing",
@@ -6159,6 +6160,7 @@ dependencies = [
"futures-core",
"pin-project-lite",
"tokio",
"tokio-util",
]
[[package]]

157
ROOT_CAUSE_ANALYSIS.md Normal file
View File

@@ -0,0 +1,157 @@
# Root Cause Analysis: Token Repetition in Streaming Text Generation
**Date:** August 27, 2025
**System:** Predict-Otron-9000 Inference Engine
**Issue:** Token repetition in streaming text generation despite successful individual token streaming implementation
## Executive Summary
The Predict-Otron-9000 system has successfully implemented individual token streaming and resolved false positive stream issues in CLI multiple invocations. However, token repetition remains a critical issue that degrades output quality. This analysis identifies the root cause as insufficient context preservation in the incremental token generation process, particularly for Gemma model variants.
## Technical Background
### System Architecture
The Predict-Otron-9000 consists of several key components:
1. **Inference Engine** (`crates/inference-engine/`): Core text generation logic
- `TokenOutputStream`: Handles token-by-token decoding and streaming
- `TextGeneration`: Main generation logic with streaming support
- `server.rs`: HTTP API with Server-Sent Events (SSE) streaming
2. **CLI Client** (`cli.ts`): TypeScript client for interacting with the inference engine
3. **Model Support**: Gemma-1, Gemma-2, and Gemma-3 model variants
### Streaming Implementation Changes
#### Individual Token Generation ✅ RESOLVED
**Previous Behavior:** Tokens were generated in batches and sent all at once.
**Current Implementation:**
- `TokenOutputStream.next_token()` processes individual tokens with incremental decoding
- Modified to "include all tokens, not just alphanumeric ones" (token_output_stream.rs:44)
- Server streams tokens via SSE using callback mechanism in `TextGeneration.run_with_streaming()`
#### CLI Multiple Invocation Support ✅ RESOLVED
**Previous Issue:** Multiple CLI invocations received false positive streams from previous sessions.
**Current Solution:**
- Each CLI invocation creates a fresh OpenAI client connection
- Server calls `text_gen.reset_state()` before each streaming request
- `TokenOutputStream.clear()` resets token buffers and indices
- Penalty cache is cleared for each new generation
## Root Cause Analysis: Token Repetition
### Primary Root Cause: Insufficient Context Window
The token repetition issue stems from **severe context limitation** in the incremental generation process:
#### 1. Gemma Model Special Handling (Lines 694-806 in text_generation.rs)
```rust
// Use just the last token for subsequent iterations to avoid shape mismatch
let context_tokens = &tokens[(tokens.len()-1)..];
let start_pos = tokens.len() - 1;
```
**Problem:** For Gemma-2 and Gemma-3 models, only the **last single token** is used for subsequent forward passes. This eliminates virtually all context, forcing the model to generate based on minimal information.
#### 2. Standard Model Handling (Lines 808-850 in text_generation.rs)
```rust
let context_size = if index > 0 { 1 } else { tokens.len() };
let start_pos = tokens.len().saturating_sub(context_size);
let ctxt = &tokens[start_pos..];
```
**Problem:** After the first token, context is limited to just **1 token** for all subsequent generations, again severely restricting the model's ability to maintain coherent context.
#### 3. Penalty Cache Clearing
```rust
// Clear penalty cache for new generation
self.penalty_cache.clear();
```
**Contributing Factor:** The repeat penalty cache is cleared at the start of each streaming generation, reducing the effectiveness of repetition prevention mechanisms.
### Secondary Contributing Factors
1. **Shape Compatibility Workaround**: The single-token context approach was implemented to "avoid shape mismatch" in Gemma models, prioritizing technical compatibility over generation quality.
2. **Incremental Context Loss**: Each forward pass operates with minimal historical context, making it impossible for the model to understand what it has already generated.
3. **Inadequate Repeat Penalty Context**: The repeat penalty mechanism (`apply_cached_repeat_penalty`) has limited effectiveness when working with truncated context windows.
## Impact Analysis
### Performance Impact
- **Positive**: Individual token streaming provides responsive user experience
- **Positive**: CLI multiple invocations work correctly without interference
- **Negative**: Poor output quality due to repetitive content
### User Experience Impact
- **Critical**: Generated text contains significant repetition, reducing practical utility
- **Positive**: Real-time streaming provides immediate feedback
- **Positive**: Consistent behavior across multiple CLI sessions
### Technical Debt
- **High**: Current implementation prioritizes technical workarounds over generation quality
- **Medium**: Context limitation approach creates maintenance burden
- **Low**: Streaming infrastructure is well-architected and maintainable
## Timeline and Change History
Based on code analysis, the following changes were implemented:
1. **Token Streaming Enhancement**: Modified `TokenOutputStream` to include all tokens, not just alphanumeric
2. **Individual Token Callbacks**: Implemented streaming callbacks in `TextGeneration.run_with_streaming()`
3. **CLI State Management**: Added proper state reset and fresh connections
4. **Context Limitation Implementation**: Applied single-token context for incremental generation
5. **SSE Integration**: Implemented Server-Sent Events for real-time token delivery
## Recommendations for Future Iterations
### Immediate Priority (Critical)
1. **Implement Sliding Window Context**: Replace single-token context with a configurable sliding window (e.g., last 50-100 tokens)
2. **Context-Aware Repeat Penalty**: Maintain repeat penalty context across the full generation window
3. **Model-Specific Context Handling**: Develop proper context management for each Gemma variant without sacrificing context size
### Medium-Term Improvements
1. **Dynamic Context Sizing**: Implement adaptive context windows based on available memory and model capabilities
2. **Advanced Repetition Detection**: Implement semantic-level repetition detection beyond token-level penalties
3. **Context Compression**: Explore context compression techniques to maintain longer effective context windows
### Long-Term Enhancements
1. **Beam Search Integration**: Implement beam search with streaming for better output quality
2. **Adaptive Sampling**: Dynamic adjustment of sampling parameters based on repetition detection
3. **Model Fine-tuning**: Consider fine-tuning approaches to reduce repetition tendency at the model level
## Monitoring and Validation
### Key Metrics to Track
1. **Repetition Rate**: Measure token and n-gram repetition frequency
2. **Context Utilization**: Monitor effective context window usage
3. **Generation Quality**: Track coherence and diversity metrics
4. **Streaming Performance**: Maintain current responsiveness standards
### Testing Strategy
1. **Repetition Benchmarks**: Create standardized tests for repetition detection
2. **Context Window Testing**: Validate context preservation across different window sizes
3. **Model Variant Testing**: Ensure consistent behavior across Gemma-1, Gemma-2, and Gemma-3
4. **Regression Testing**: Maintain streaming functionality during context improvements
## Conclusion
The Predict-Otron-9000 has successfully achieved individual token streaming and eliminated false positive streams in CLI usage. However, the current implementation's approach to context management—using only single tokens for incremental generation—is the primary root cause of token repetition issues.
The solution requires balancing technical compatibility with generation quality by implementing proper sliding window context management while maintaining the current streaming performance and reliability. This represents a critical technical debt that should be addressed in the next development iteration to realize the system's full potential.
**Priority Level:** Critical
**Complexity:** Medium
**Risk Level:** Low (improvements can be made incrementally)
**User Impact:** High (significant quality improvement expected)

252
cli.ts
View File

@@ -3,8 +3,30 @@
import OpenAI from "openai";
import { parseArgs } from "util";
// =====================
// Config
// =====================
const DEFAULT_MODEL = "gemma-3-1b-it";
const DEFAULT_MAX_TOKENS = 100;
const DEFAULT_MAX_TOKENS = 256;
// Toggle this to reduce log overhead during timing runs
const PRINT_CHUNK_DEBUG = false;
// How many rows to show in the timing tables
const SHOW_FIRST_N = 3;
const SHOW_SLOWEST_N = 3;
// =====================
// Helpers
// =====================
const now = () => performance.now();
type ChunkStat = {
index: number;
tSinceRequestStartMs: number;
dtSincePrevMs: number;
contentChars: number;
};
function printHelp() {
console.log(`
@@ -30,20 +52,12 @@ Start it with: ./run_server.sh
}
const { values, positionals } = parseArgs({
args: Bun.argv,
args: process.argv.slice(2),
options: {
model: {
type: 'string',
},
prompt: {
type: 'string',
},
help: {
type: 'boolean',
},
'list-models': {
type: 'boolean',
},
model: { type: "string" },
prompt: { type: "string" },
help: { type: "boolean" },
"list-models": { type: "boolean" },
},
strict: false,
allowPositionals: true,
@@ -55,16 +69,20 @@ async function requestLocalOpenAI(model: string, userPrompt: string) {
apiKey: "not used",
});
try {
console.log("[DEBUG] Creating chat completion request...");
return openai.chat.completions.create({
model: model,
model,
max_tokens: DEFAULT_MAX_TOKENS,
stream: true,
messages: [
{name: "assistant_1", role: "system", content: "I am a helpful assistant" },
{name: "user_1", role: "user", content: userPrompt}
]
{
role: "system",
content: "You are a helpful assistant who responds thoughtfully and concisely.",
},
{ role: "user", content: userPrompt },
],
});
} catch (e) {
} catch (e: any) {
console.error("[ERROR] Failed to connect to local OpenAI server:", e.message);
console.error("[HINT] Make sure the server is running at http://localhost:8080");
console.error("[HINT] Start it with: ./run_server.sh");
@@ -81,7 +99,7 @@ async function listModels() {
const models = await openai.models.list();
console.log(`[INFO] Available models from http://localhost:8080/v1:`);
console.log("---");
if (models.data && models.data.length > 0) {
models.data.forEach((model, index) => {
console.log(`${index + 1}. ${model.id}`);
@@ -93,8 +111,7 @@ async function listModels() {
} else {
console.log("No models found.");
}
} catch (e) {
} catch (e: any) {
console.error("[ERROR] Failed to fetch models from local OpenAI server:", e.message);
console.error("[HINT] Make sure the server is running at http://localhost:8080");
console.error("[HINT] Start it with: ./run_server.sh");
@@ -102,34 +119,58 @@ async function listModels() {
}
}
// =====================
// Timing math
// =====================
function median(nums: number[]) {
if (nums.length === 0) return 0;
const s = [...nums].sort((a, b) => a - b);
const mid = Math.floor(s.length / 2);
return s.length % 2 ? s[mid] : (s[mid - 1] + s[mid]) / 2;
}
function quantile(nums: number[], q: number) {
if (nums.length === 0) return 0;
const s = [...nums].sort((a, b) => a - b);
const pos = (s.length - 1) * q;
const base = Math.floor(pos);
const rest = pos - base;
return s[base + 1] !== undefined ? s[base] + rest * (s[base + 1] - s[base]) : s[base];
}
function ms(n: number) {
return `${n.toFixed(1)} ms`;
}
// =====================
// Main
// =====================
async function main() {
// Show help if requested
const tProgramStart = now();
if (values.help) {
printHelp();
process.exit(0);
}
// List models if requested
if (values['list-models']) {
if (values["list-models"]) {
try {
await listModels();
process.exit(0);
} catch (error) {
} catch (error: any) {
console.error("\n[ERROR] Failed to list models:", error.message);
process.exit(1);
}
}
// Get the prompt from either --prompt flag or positional argument
const prompt = values.prompt || positionals[2]; // positionals[0] is 'bun', positionals[1] is 'client_cli.ts'
const prompt = values.prompt ?? positionals[0];
if (!prompt) {
console.error("[ERROR] No prompt provided!");
printHelp();
process.exit(1);
}
// Get the model (use default if not provided)
const model = values.model || DEFAULT_MODEL;
console.log(`[INFO] Using model: ${model}`);
@@ -137,32 +178,163 @@ async function main() {
console.log(`[INFO] Connecting to: http://localhost:8080/v1`);
console.log("---");
const tBeforeRequest = now();
try {
console.log("[DEBUG] Initiating request to OpenAI server...");
const response = await requestLocalOpenAI(model, prompt);
// Handle streaming response
const tAfterCreate = now();
// Streaming handling + timing
let fullResponse = "";
let chunkCount = 0;
const chunkStats: ChunkStat[] = [];
let tFirstChunk: number | null = null;
let tPrevChunk: number | null = null;
console.log("[INFO] Waiting for model to generate response...");
let loadingInterval;
if (!PRINT_CHUNK_DEBUG) {
// Show loading animation only if not in debug mode
const loadingChars = ['⠋', '⠙', '⠹', '⠸', '⠼', '⠴', '⠦', '⠧', '⠇', '⠏'];
let i = 0;
process.stdout.write('\r[INFO] Thinking ');
loadingInterval = setInterval(() => {
process.stdout.write(`\r[INFO] Thinking ${loadingChars[i++ % loadingChars.length]} `);
}, 80);
} else {
console.log("[DEBUG] Starting to receive streaming response...");
}
for await (const chunk of response) {
const content = chunk.choices[0]?.delta?.content;
// Clear loading animation on first chunk
if (loadingInterval) {
clearInterval(loadingInterval);
process.stdout.write('\r \r');
}
const tNow = now();
chunkCount++;
// Extract content (delta) if present
const content = chunk.choices?.[0]?.delta?.content ?? "";
if (PRINT_CHUNK_DEBUG) {
console.log(`[DEBUG] Received chunk #${chunkCount}:`, JSON.stringify(chunk));
if (content) console.log(`[DEBUG] Chunk content: "${content}"`);
}
if (content) {
process.stdout.write(content);
fullResponse += content;
}
if (tFirstChunk === null) tFirstChunk = tNow;
const dtSincePrev = tPrevChunk === null ? 0 : tNow - tPrevChunk;
chunkStats.push({
index: chunkCount,
tSinceRequestStartMs: tNow - tBeforeRequest,
dtSincePrevMs: dtSincePrev,
contentChars: content.length,
});
tPrevChunk = tNow;
}
// =========
// Summary
// =========
const tStreamEnd = now();
const totalChars = fullResponse.length;
console.log("\n---");
console.log(`[INFO] Response completed. Total length: ${fullResponse.length} characters`);
} catch (error) {
console.log(`[DEBUG] Stream completed after ${chunkCount} chunks`);
console.log(`[INFO] Response completed. Total length: ${totalChars} characters`);
// Build timing metrics
const ttfbMs = (tFirstChunk ?? tStreamEnd) - tAfterCreate; // time from create() resolved → first chunk
const createOverheadMs = tAfterCreate - tBeforeRequest; // time spent awaiting create() promise
const totalSinceRequestMs = tStreamEnd - tBeforeRequest; // from just before create() to last chunk
const streamDurationMs =
tFirstChunk === null ? 0 : tStreamEnd - tFirstChunk;
const gaps = chunkStats
.map((c) => c.dtSincePrevMs)
// ignore the first "gap" which is 0 by construction
.slice(1);
const avgGapMs = gaps.length ? gaps.reduce((a, b) => a + b, 0) / gaps.length : 0;
const medGapMs = median(gaps);
const p95GapMs = quantile(gaps, 0.95);
let maxGapMs = 0;
let maxGapAtChunk = 0;
for (let i = 0; i < gaps.length; i++) {
if (gaps[i] > maxGapMs) {
maxGapMs = gaps[i];
maxGapAtChunk = i + 2; // +1 to move from 0-based, +1 because we sliced starting at second chunk
}
}
// Pretty print summary
console.log("\n=== Timing Summary ===");
console.log(`create() await time: ${ms(createOverheadMs)}`);
console.log(`TTFB (to 1st chunk): ${ms(ttfbMs)}`);
console.log(`Stream duration: ${ms(streamDurationMs)}`);
console.log(`End-to-end (req→last): ${ms(totalSinceRequestMs)}`);
console.log(`Chunks: ${chunkCount}`);
console.log(`Total content chars: ${totalChars}`);
console.log(`Avg chars/chunk: ${(chunkCount ? totalChars / chunkCount : 0).toFixed(1)}`);
console.log(`Inter-chunk gap (avg): ${ms(avgGapMs)}`);
console.log(`Inter-chunk gap (median): ${ms(medGapMs)}`);
console.log(`Inter-chunk gap (p95): ${ms(p95GapMs)}`);
if (gaps.length > 0) {
console.log(`Largest gap: ${ms(maxGapMs)} (before chunk #${maxGapAtChunk})`);
}
// Small tables: first N and slowest N gaps
const firstRows = chunkStats.slice(0, SHOW_FIRST_N).map((c) => ({
chunk: c.index,
"t since request": `${c.tSinceRequestStartMs.toFixed(1)} ms`,
"dt since prev": `${c.dtSincePrevMs.toFixed(1)} ms`,
"chars": c.contentChars,
}));
const slowestRows = chunkStats
.slice(1) // skip first (no meaningful gap)
.sort((a, b) => b.dtSincePrevMs - a.dtSincePrevMs)
.slice(0, SHOW_SLOWEST_N)
.map((c) => ({
chunk: c.index,
"t since request": `${c.tSinceRequestStartMs.toFixed(1)} ms`,
"dt since prev": `${c.dtSincePrevMs.toFixed(1)} ms`,
"chars": c.contentChars,
}));
if (firstRows.length > 0) {
console.log("\n--- First chunk timings ---");
// @ts-ignore Bun/Node support console.table
console.table(firstRows);
}
if (slowestRows.length > 0) {
console.log(`\n--- Slowest ${SHOW_SLOWEST_N} gaps ---`);
// @ts-ignore
console.table(slowestRows);
}
const tProgramEnd = now();
console.log("\n=== Program Overhead ===");
console.log(`Total program runtime: ${ms(tProgramEnd - tProgramStart)}`);
} catch (error: any) {
console.error("\n[ERROR] Request failed:", error.message);
process.exit(1);
}
}
// Run the main function
main().catch(error => {
main().catch((error) => {
console.error("[FATAL ERROR]:", error);
process.exit(1);
});

9
crates/embeddings-engine/Cargo.toml
View File

@@ -24,3 +24,12 @@ tracing-subscriber = { version = "0.3", features = ["env-filter"] }
rand = "0.8.5"
async-openai = "0.28.3"
once_cell = "1.19.0"
[package.metadata.kube]
image = "ghcr.io/geoffsee/embeddings-service:latest"
replicas = 1
port = 8080
resources.cpu = "500m"
resources.memory = "256Mi"
#ingress.host = "my-service.example.com"
#env = { RUST_LOG = "info", DATABASE_URL = "postgres://..." }

2
crates/embeddings-engine/src/main.rs
View File

@@ -10,7 +10,7 @@ use std::env;
use tower_http::trace::TraceLayer;
use tracing;
const DEFAULT_SERVER_HOST: &str = "0.0.0.0";
const DEFAULT_SERVER_HOST: &str = "127.0.0.1";
const DEFAULT_SERVER_PORT: &str = "8080";
async fn root() -> &'static str {

12
crates/inference-engine/Cargo.toml
View File

@@ -44,6 +44,7 @@ axum = { version = "0.8.4", features = ["json"] }
tower = "0.5.2"
tower-http = { version = "0.6.6", features = ["cors"] }
tokio = { version = "1.43.0", features = ["full"] }
tokio-stream = { version = "0.1.16", features = ["sync"] }
either = { version = "1.9.0", features = ["serde"] }
utoipa = { version = "4.2.0", features = ["axum_extras"] }
uuid = { version = "1.7.0", features = ["v4"] }
@@ -80,4 +81,13 @@ tokio = "1.43.0"
[build-dependencies]
anyhow = { version = "1", features = ["backtrace"] }
bindgen_cuda = { version = "0.1.1", optional = true }
bindgen_cuda = { version = "0.1.1", optional = true }
[package.metadata.kube]
image = "ghcr.io/geoffsee/inference-service:latest"
replicas = 1
port = 8080
resources.cpu = "500m"
resources.memory = "256Mi"
#ingress.host = "my-service.example.com"
#env = { RUST_LOG = "info", DATABASE_URL = "postgres://..." }

2
crates/inference-engine/src/lib.rs
View File

@@ -17,7 +17,7 @@ use std::env;
use tracing_subscriber::{layer::SubscriberExt, util::SubscriberInitExt};
/// Server configuration constants
pub const DEFAULT_SERVER_HOST: &str = "0.0.0.0";
pub const DEFAULT_SERVER_HOST: &str = "127.0.0.1";
pub const DEFAULT_SERVER_PORT: &str = "8080";
/// Get server configuration from environment variables with defaults

733
crates/inference-engine/src/server.rs
View File

@@ -5,14 +5,13 @@ use axum::{
routing::{get, post},
Json, Router,
};
use futures_util::stream::{self, Stream};
use std::convert::Infallible;
use candle_core::DType;
use candle_nn::VarBuilder;
use futures_util::stream::{self, Stream};
use tokio_stream::wrappers::UnboundedReceiverStream;
use std::convert::Infallible;
use std::{path::PathBuf, sync::Arc};
use std::time::Duration;
use tokio::sync::Mutex;
use tokio::time;
use tokio::sync::{Mutex, mpsc};
use tower_http::cors::{Any, CorsLayer};
use uuid::Uuid;
@@ -46,40 +45,26 @@ impl Default for AppState {
let text_generation = build_pipeline(args.clone());
Self {
text_generation: Arc::new(Mutex::new(text_generation)),
model_id: String::new(),
model_id: args.model_id.clone(),
build_args: args,
}
}
}
// -------------------------
// Pipeline configuration
// -------------------------
#[derive(Debug, Clone)]
pub struct PipelineArgs {
/// HF model repo id, e.g. "google/gemma-2b"
pub model_id: String,
/// Which internal model family to instantiate
pub which: Which,
/// Optional HF revision/branch/tag; None => "main"
pub revision: Option<String>,
/// Optional explicit tokenizer path
pub tokenizer_path: Option<PathBuf>,
/// Optional explicit config path
pub config_path: Option<PathBuf>,
/// Optional explicit weight paths. If empty, they will be resolved from the hub.
pub weight_paths: Vec<PathBuf>,
/// Runtime toggles
pub use_flash_attn: bool,
pub force_cpu: bool,
/// Sampling / decoding params
pub seed: u64,
pub temperature: Option<f64>,
pub top_p: Option<f64>,
@@ -98,39 +83,43 @@ impl Default for PipelineArgs {
weight_paths: Vec::new(),
use_flash_attn: false,
force_cpu: false,
seed: 0,
temperature: None,
top_p: None,
repeat_penalty: 0.0,
repeat_last_n: 0,
seed: 299792458, // Speed of light in vacuum (m/s)
temperature: Some(0.8), // Good balance between creativity and coherence
top_p: Some(0.9), // Keep diverse but reasonable options
repeat_penalty: 1.2, // Stronger penalty for repetition to prevent looping
repeat_last_n: 64, // Consider last 64 tokens for repetition
}
}
}
// If no owner/org is present, prefix with a sensible default (tweak as you like).
fn normalize_model_id(model_id: &str) -> String {
if model_id.contains('/') { model_id.to_string() } else { format!("google/{}", model_id) }
if model_id.contains('/') {
model_id.to_string()
} else {
format!("google/{}", model_id)
}
}
// Quick existence check, mapping 404 into a helpful message.
fn ensure_repo_exists(api: &Api, model_id: &str, revision: &str) -> anyhow::Result<()> {
let repo = api.repo(Repo::with_revision(model_id.to_string(), RepoType::Model, revision.to_string()));
let repo = api.repo(Repo::with_revision(
model_id.to_string(),
RepoType::Model,
revision.to_string(),
));
match repo.get("config.json") {
Ok(_) => Ok(()),
Err(e) => match e {
ApiError::RequestError(resp) => {
// For HF API, RequestError with 404 status is returned when repo doesn't exist
let error_str = resp.to_string();
if error_str.contains("404") {
anyhow::bail!(
"Hugging Face model repo not found: '{model_id}' at revision '{revision}'. \
Please provide a fully-qualified repo id like 'google/gemma-2b-it'."
"Hugging Face model repo not found: '{model_id}' at revision '{revision}'."
)
}
Err(anyhow::Error::new(ApiError::RequestError(resp)))
}
other => Err(anyhow::Error::new(other)),
}
},
}
}
@@ -151,18 +140,13 @@ pub fn build_pipeline(mut args: PipelineArgs) -> TextGeneration {
let api = Api::new().unwrap();
let revision = args.revision.as_deref().unwrap_or("main");
// Check if model_id is empty before normalizing it
println!("Checking model_id: '{}'", args.model_id);
println!("Trimmed model_id length: {}", args.model_id.trim().len());
if args.model_id.trim().is_empty() {
panic!("No model ID specified. Please provide a valid model ID (e.g., 'gemma-2b-it' or 'google/gemma-2b-it').");
panic!("No model ID specified.");
}
args.model_id = normalize_model_id(&args.model_id);
// Validate early (nice error if the repo/revision is wrong).
match ensure_repo_exists(&api, &args.model_id, revision) {
Ok(_) => {},
Ok(_) => {}
Err(e) => panic!("{}", e),
};
@@ -172,105 +156,82 @@ pub fn build_pipeline(mut args: PipelineArgs) -> TextGeneration {
revision.to_string(),
));
// Resolve files (prefer explicit paths; fallback to hub)
let tokenizer_path = args
.tokenizer_path
.unwrap_or_else(|| repo.get("tokenizer.json").unwrap());
let config_path = args
.config_path
.unwrap_or_else(|| repo.get("config.json").unwrap());
// Only use auto-detection if no specific model type was provided
// This ensures that explicitly specified model types are respected
if !matches!(args.which,
Which::Base2B | Which::Base7B |
Which::Instruct2B | Which::Instruct7B |
Which::InstructV1_1_2B | Which::InstructV1_1_7B |
Which::CodeBase2B | Which::CodeBase7B |
Which::CodeInstruct2B | Which::CodeInstruct7B |
Which::BaseV2_2B | Which::InstructV2_2B |
Which::BaseV2_9B | Which::InstructV2_9B |
Which::BaseV3_1B | Which::InstructV3_1B) {
// If model_id is a known value, map it directly
if !matches!(
args.which,
Which::Base2B
| Which::Base7B
| Which::Instruct2B
| Which::Instruct7B
| Which::InstructV1_1_2B
| Which::InstructV1_1_7B
| Which::CodeBase2B
| Which::CodeBase7B
| Which::CodeInstruct2B
| Which::CodeInstruct7B
| Which::BaseV2_2B
| Which::InstructV2_2B
| Which::BaseV2_9B
| Which::InstructV2_9B
| Which::BaseV3_1B
| Which::InstructV3_1B
) {
if args.model_id.contains("gemma-2-2b-it") {
args.which = Which::InstructV2_2B;
println!("Setting model type to InstructV2_2B based on model_id: {}", args.model_id);
} else if args.model_id.contains("gemma-3-1b-it") {
args.which = Which::InstructV3_1B;
println!("Setting model type to InstructV3_1B based on model_id: {}", args.model_id);
} else {
// Fallback to auto-detection from config.json
if let Ok(file) = std::fs::File::open(config_path.clone()) {
if let Ok(cfg_val) = serde_json::from_reader::<_, serde_json::Value>(file) {
if let Some(model_type) = cfg_val.get("model_type").and_then(|v| v.as_str()) {
println!("Auto-detecting model type from config.json: {}", model_type);
// Map HF model_type to an internal Which variant
if model_type.contains("gemma3") {
args.which = Which::InstructV3_1B;
println!("Setting model type to InstructV3_1B based on config");
} else if model_type.contains("gemma2") {
args.which = Which::InstructV2_2B;
println!("Setting model type to InstructV2_2B based on config");
} else {
// default to Gemma v1
args.which = Which::Instruct2B;
println!("Setting model type to Instruct2B (v1) based on config");
}
} else if let Ok(file) = std::fs::File::open(config_path.clone()) {
if let Ok(cfg_val) = serde_json::from_reader::<_, serde_json::Value>(file) {
if let Some(model_type) = cfg_val.get("model_type").and_then(|v| v.as_str()) {
if model_type.contains("gemma3") {
args.which = Which::InstructV3_1B;
} else if model_type.contains("gemma2") {
args.which = Which::InstructV2_2B;
} else {
args.which = Which::Instruct2B;
}
}
}
}
} else {
println!("Using explicitly specified model type: {:?}", args.which);
}
// Resolve weight files: try a single-file first, then fall back to sharded index
let weight_paths = if !args.weight_paths.is_empty() {
args.weight_paths
} else {
match repo.get("model.safetensors") {
Ok(single) => vec![single],
Err(_) => {
match utilities_lib::hub_load_safetensors(&repo, "model.safetensors.index.json") {
Ok(paths) => paths,
Err(e) => {
panic!(
"Unable to locate model weights for '{}'. Tried 'model.safetensors' and 'model.safetensors.index.json'. Underlying error: {}",
args.model_id, e
);
}
Err(_) => match utilities_lib::hub_load_safetensors(&repo, "model.safetensors.index.json") {
Ok(paths) => paths,
Err(e) => {
panic!("Unable to locate model weights: {}", e);
}
}
},
}
};
println!("retrieved the files in {:?}", start.elapsed());
let tokenizer = Tokenizer::from_file(tokenizer_path)
.map_err(anyhow::Error::msg)
.unwrap();
let start = std::time::Instant::now();
let tokenizer = Tokenizer::from_file(tokenizer_path).unwrap();
let initial_device = utilities_lib::device(args.force_cpu).unwrap();
// Check if we're using a V3 model (Gemma 3) and if we're on Metal (macOS)
let is_v3_model = args.which.is_v3_model();
let is_metal = !initial_device.is_cpu() && candle_core::utils::metal_is_available() && !args.force_cpu;
// Use CPU for V3 models on Metal due to missing implementations
let is_metal = !initial_device.is_cpu()
&& candle_core::utils::metal_is_available()
&& !args.force_cpu;
let device = if is_v3_model && is_metal {
println!("Note: Using CPU for Gemma 3 model due to missing Metal implementations for required operations (e.g., rotary-emb).");
candle_core::Device::Cpu
} else {
initial_device
};
let dtype = if device.is_cuda() { DType::BF16 } else { DType::F32 };
// Keep original device + dtype
let dtype = if device.is_cuda() { DType::BF16 } else { DType::F32 };
let vb = unsafe { VarBuilder::from_mmaped_safetensors(&weight_paths, dtype, &device).unwrap() };
let model = match args.which {
@@ -285,23 +246,18 @@ pub fn build_pipeline(mut args: PipelineArgs) -> TextGeneration {
| Which::CodeInstruct2B
| Which::CodeInstruct7B => {
let config: Config1 = serde_json::from_reader(std::fs::File::open(config_path.clone()).unwrap()).unwrap();
let model = Model1::new(args.use_flash_attn, &config, vb).unwrap();
GemmaModel::V1(model)
GemmaModel::V1(Model1::new(args.use_flash_attn, &config, vb).unwrap())
}
Which::BaseV2_2B | Which::InstructV2_2B | Which::BaseV2_9B | Which::InstructV2_9B => {
let config: Config2 = serde_json::from_reader(std::fs::File::open(config_path.clone()).unwrap()).unwrap();
let model = Model2::new(args.use_flash_attn, &config, vb).unwrap();
GemmaModel::V2(model)
GemmaModel::V2(Model2::new(args.use_flash_attn, &config, vb).unwrap())
}
Which::BaseV3_1B | Which::InstructV3_1B => {
let config: Config3 = serde_json::from_reader(std::fs::File::open(config_path).unwrap()).unwrap();
let model = Model3::new(args.use_flash_attn, &config, vb).unwrap();
GemmaModel::V3(model)
GemmaModel::V3(Model3::new(args.use_flash_attn, &config, vb).unwrap())
}
};
println!("loaded the model in {:?}", start.elapsed());
TextGeneration::new(
model,
tokenizer,
@@ -314,6 +270,43 @@ pub fn build_pipeline(mut args: PipelineArgs) -> TextGeneration {
)
}
fn build_gemma_prompt(messages: &[Message]) -> String {
let mut prompt = String::new();
let mut system_prompt: Option<String> = None;
for message in messages {
let content = match &message.content {
Some(content) => match &content.0 {
Either::Left(text) => text.clone(),
Either::Right(_) => "".to_string(),
},
None => "".to_string(),
};
match message.role.as_str() {
"system" => system_prompt = Some(content),
"user" => {
prompt.push_str("<start_of_turn>user\n");
if let Some(sys_prompt) = system_prompt.take() {
prompt.push_str(&sys_prompt);
prompt.push_str("\n\n");
}
prompt.push_str(&content);
prompt.push_str("<end_of_turn>\n");
}
"assistant" => {
prompt.push_str("<start_of_turn>model\n");
prompt.push_str(&content);
prompt.push_str("<end_of_turn>\n");
}
_ => {}
}
}
prompt.push_str("<start_of_turn>model\n");
prompt
}
// -------------------------
// OpenAI-compatible handler
// -------------------------
@@ -322,72 +315,68 @@ pub async fn chat_completions(
State(state): State<AppState>,
Json(request): Json<ChatCompletionRequest>,
) -> Result<impl IntoResponse, (StatusCode, String)> {
// If streaming was requested, this function shouldn't be called
// A separate route handles streaming requests
if !request.stream.unwrap_or(false) {
return Ok(chat_completions_non_streaming_proxy(state, request).await.into_response())
return Ok(chat_completions_non_streaming_proxy(state, request).await.into_response());
}
Ok(chat_completions_stream(state, request).await.into_response())
}
pub async fn chat_completions_non_streaming_proxy(state: AppState, request: ChatCompletionRequest) -> Result<impl IntoResponse, (StatusCode, Json<Value>)> {
// Non-streaming response - original implementation
let mut prompt = String::new();
pub async fn chat_completions_non_streaming_proxy(
state: AppState,
request: ChatCompletionRequest,
) -> Result<impl IntoResponse, (StatusCode, Json<Value>)> {
let prompt = build_gemma_prompt(&request.messages);
// Convert messages to a prompt string
for message in &request.messages {
let role = &message.role;
let content = match &message.content {
Some(content) => match &content.0 {
Either::Left(text) => text.clone(),
Either::Right(_) => "".to_string(), // Handle complex content if needed
},
None => "".to_string(),
};
match role.as_str() {
"system" => prompt.push_str(&format!("System: {}\n", content)),
"user" => prompt.push_str(&format!("User: {}\n", content)),
"assistant" => prompt.push_str(&format!("Assistant: {}\n", content)),
_ => prompt.push_str(&format!("{}: {}\n", role, content)),
}
}
prompt.push_str("Assistant: ");
let model_id = state.model_id.clone();
// Generate
let mut output = Vec::new();
{
// Recreate TextGeneration instance to ensure completely fresh state
// This prevents KV cache persistence that causes tensor shape mismatches
let fresh_text_gen = build_pipeline(state.build_args.clone());
let mut text_gen = state.text_generation.lock().await;
*text_gen = fresh_text_gen;
let mut buffer = Vec::new();
let max_tokens = request.max_tokens.unwrap_or(1000);
let result = text_gen.run_with_output(&prompt, max_tokens, &mut buffer);
if let Err(e) = result {
// Enforce model selection behavior: reject if a different model is requested
let configured_model = state.build_args.model_id.clone();
let requested_model = request.model.clone();
if requested_model.to_lowercase() != "default" {
let normalized_requested = normalize_model_id(&requested_model);
if normalized_requested != configured_model {
return Err((
StatusCode::BAD_REQUEST,
Json(serde_json::json!({
"error": {
"message": format!("Error generating text: {}", e),
"type": "text_generation_error"
"message": format!(
"Requested model '{}' is not available. This server is running '{}' only.",
requested_model, configured_model
),
"type": "model_mismatch"
}
})),
));
}
}
if let Ok(text) = String::from_utf8(buffer) {
output.push(text);
let model_id = state.model_id.clone();
let mut buffer = Vec::new();
{
let mut text_gen = state.text_generation.lock().await;
// Reset per-request state without rebuilding the whole pipeline
text_gen.reset_state();
let max_tokens = request.max_tokens.unwrap_or(1000);
if let Err(e) = text_gen.run_with_output(&prompt, max_tokens, &mut buffer) {
return Err((
StatusCode::BAD_REQUEST,
Json(serde_json::json!({
"error": { "message": format!("Error generating text: {}", e) }
})),
));
}
}
let completion = output.join("");
let completion = match String::from_utf8(buffer) {
Ok(s) => s,
Err(e) => {
return Err((
StatusCode::INTERNAL_SERVER_ERROR,
Json(serde_json::json!({
"error": { "message": format!("UTF-8 conversion error: {}", e) }
})),
));
}
};
let response = ChatCompletionResponse {
id: format!("chatcmpl-{}", Uuid::new_v4().to_string().replace('-', "")),
@@ -407,7 +396,6 @@ pub async fn chat_completions_non_streaming_proxy(state: AppState, request: Chat
finish_reason: "stop".to_string(),
}],
usage: Usage {
// still rough estimates
prompt_tokens: prompt.len() / 4,
completion_tokens: completion.len() / 4,
total_tokens: (prompt.len() + completion.len()) / 4,
@@ -415,185 +403,195 @@ pub async fn chat_completions_non_streaming_proxy(state: AppState, request: Chat
};
Ok(Json(response).into_response())
}
// -------------------------
// Streaming implementation
// -------------------------
pub async fn chat_completions_stream(
state: AppState,
chat_completion_request: ChatCompletionRequest,
) -> Result<Sse<impl Stream<Item = Result<Event, Infallible>>>, (StatusCode, Json<serde_json::Value>)> {
// Call the handler function
handle_streaming_request(state, chat_completion_request).await
request: ChatCompletionRequest,
) -> Result<Sse<impl Stream<Item = Result<Event, Infallible>>>, (StatusCode, Json<Value>)> {
handle_streaming_request(state, request).await
}
/// Handle streaming requests with Server-Sent Events (SSE)
async fn handle_streaming_request(
state: AppState,
request: ChatCompletionRequest
) -> Result<Sse<impl Stream<Item = Result<Event, Infallible>>>, (StatusCode, Json<serde_json::Value>)> {
// Generate a unique ID for this completion
state: AppState,
request: ChatCompletionRequest,
) -> Result<Sse<impl Stream<Item = Result<Event, Infallible>>>, (StatusCode, Json<Value>)> {
// Validate requested model vs configured model
let configured_model = state.build_args.model_id.clone();
let requested_model = request.model.clone();
if requested_model.to_lowercase() != "default" {
let normalized_requested = normalize_model_id(&requested_model);
if normalized_requested != configured_model {
return Err((
StatusCode::BAD_REQUEST,
Json(serde_json::json!({
"error": {
"message": format!(
"Requested model '{}' is not available. This server is running '{}' only.",
requested_model, configured_model
),
"type": "model_mismatch"
}
})),
));
}
}
// Generate a unique ID and metadata
let response_id = format!("chatcmpl-{}", Uuid::new_v4().to_string().replace('-', ""));
let created = std::time::SystemTime::now()
.duration_since(std::time::UNIX_EPOCH)
.unwrap_or_default()
.as_secs();
let model_id = state.model_id.clone();
// Convert messages to a prompt string (same as non-streaming)
let mut prompt = String::new();
for message in &request.messages {
let role = &message.role;
let content = match &message.content {
Some(content) => match &content.0 {
Either::Left(text) => text.clone(),
Either::Right(_) => "".to_string(), // Handle complex content if needed
},
None => "".to_string(),
};
match role.as_str() {
"system" => prompt.push_str(&format!("System: {}\n", content)),
"user" => prompt.push_str(&format!("User: {}\n", content)),
"assistant" => prompt.push_str(&format!("Assistant: {}\n", content)),
_ => prompt.push_str(&format!("{}: {}\n", role, content)),
}
// Build prompt
let prompt = build_gemma_prompt(&request.messages);
tracing::debug!("Formatted prompt: {}", prompt);
// Channel for streaming SSE events
let (tx, rx) = mpsc::unbounded_channel::<Result<Event, Infallible>>();
// Send initial role event
let initial_chunk = ChatCompletionChunk {
id: response_id.clone(),
object: "chat.completion.chunk".to_string(),
created,
model: model_id.clone(),
choices: vec![ChatCompletionChunkChoice {
index: 0,
delta: Delta { role: Some("assistant".to_string()), content: None },
finish_reason: None,
}],
};
if let Ok(json) = serde_json::to_string(&initial_chunk) {
let _ = tx.send(Ok(Event::default().data(json)));
}
prompt.push_str("Assistant: ");
// Generate text using existing buffer-based approach
let mut buffer = Vec::new();
{
// Recreate TextGeneration instance to ensure completely fresh state
// This prevents KV cache persistence that causes tensor shape mismatches
let fresh_text_gen = build_pipeline(state.build_args.clone());
let mut text_gen = state.text_generation.lock().await;
*text_gen = fresh_text_gen;
// Spawn generation task that streams tokens as they are generated
let state_clone = state.clone();
let response_id_clone = response_id.clone();
tokio::spawn(async move {
let max_tokens = request.max_tokens.unwrap_or(1000);
let mut text_gen = state_clone.text_generation.lock().await;
text_gen.reset_state();
// Stream tokens via callback with repetition detection
let mut recent_tokens = Vec::new();
let mut repetition_count = 0;
const MAX_REPETITION_COUNT: usize = 5; // Stop after 5 consecutive repetitions
const REPETITION_WINDOW: usize = 8; // Look at last 8 tokens for patterns
if let Err(e) = text_gen.run_with_output(&prompt, max_tokens, &mut buffer) {
return Err((
StatusCode::BAD_REQUEST,
Json(serde_json::json!({
"error": {
"message": format!("Error generating text: {}", e),
"type": "text_generation_error"
let result = text_gen.run_with_streaming(&prompt, max_tokens, |token| {
// Debug log to verify token content
tracing::debug!("Streaming token: '{}'", token);
// Skip sending empty tokens
if token.is_empty() {
tracing::debug!("Skipping empty token");
return Ok(());
}
// Add token to recent history for repetition detection
recent_tokens.push(token.to_string());
if recent_tokens.len() > REPETITION_WINDOW {
recent_tokens.remove(0);
}
// Check for repetitive patterns
if recent_tokens.len() >= 4 {
let last_token = &recent_tokens[recent_tokens.len() - 1];
let second_last = &recent_tokens[recent_tokens.len() - 2];
// Check if we're repeating the same token or pattern
if last_token == second_last ||
(last_token.trim() == "plus" && second_last.trim() == "plus") ||
(recent_tokens.len() >= 6 &&
recent_tokens[recent_tokens.len()-3..].iter().all(|t| t.trim() == "plus" || t.trim().is_empty())) {
repetition_count += 1;
tracing::warn!("Detected repetition pattern: '{}' (count: {})", last_token, repetition_count);
if repetition_count >= MAX_REPETITION_COUNT {
tracing::info!("Stopping generation due to excessive repetition");
return Err(anyhow::Error::msg("Repetition detected - stopping generation"));
}
})),
));
}
}
// Convert buffer to string
let generated_text = match String::from_utf8(buffer) {
Ok(text) => text,
Err(e) => {
return Err((
StatusCode::INTERNAL_SERVER_ERROR,
Json(serde_json::json!({
"error": {
"message": format!("Error converting generated text to UTF-8: {}", e),
"type": "encoding_error"
}
})),
));
}
};
tracing::debug!("Generated text for streaming: {}", generated_text);
// Split the generated text into chunks for streaming
// This is a simplified approach - ideally we'd use proper tokenization
let chunks: Vec<String> = if !generated_text.is_empty() {
// Split by words for more natural streaming (simple approach)
generated_text.split_whitespace()
.map(|word| word.to_string() + " ")
.collect()
} else {
// If no text was generated, provide a default response
vec!["Abraham Lincoln was the 16th president of the United States.".to_string()]
};
// Create a vector to hold all the events (both chunks and DONE)
let mut events = Vec::new();
// First event includes the role
if !chunks.is_empty() {
let first_chunk = &chunks[0];
let chunk = ChatCompletionChunk {
id: response_id.clone(),
object: "chat.completion.chunk".to_string(),
created,
model: model_id.clone(),
choices: vec![ChatCompletionChunkChoice {
index: 0,
delta: Delta {
role: Some("assistant".to_string()),
content: Some(first_chunk.clone()),
},
finish_reason: None,
}],
};
if let Ok(json) = serde_json::to_string(&chunk) {
events.push(Ok(Event::default().data(json)));
}
// Add remaining chunks
for chunk_text in chunks.iter().skip(1) {
} else {
repetition_count = 0; // Reset counter if pattern breaks
}
}
let chunk = ChatCompletionChunk {
id: response_id.clone(),
id: response_id_clone.clone(),
object: "chat.completion.chunk".to_string(),
created,
model: model_id.clone(),
choices: vec![ChatCompletionChunkChoice {
index: 0,
delta: Delta {
role: None,
content: Some(chunk_text.clone()),
},
delta: Delta { role: None, content: Some(token.to_string()) },
finish_reason: None,
}],
};
if let Ok(json) = serde_json::to_string(&chunk) {
events.push(Ok(Event::default().data(json)));
tracing::debug!("Sending chunk with content: '{}'", token);
let _ = tx.send(Ok(Event::default().data(json)));
}
}
Ok(())
}).await;
// Add final chunk with finish_reason
// Log result of generation
match result {
Ok(_) => tracing::debug!("Text generation completed successfully"),
Err(e) => tracing::info!("Text generation stopped: {}", e),
}
// Send final stop chunk and DONE marker
let final_chunk = ChatCompletionChunk {
id: response_id,
id: response_id_clone.clone(),
object: "chat.completion.chunk".to_string(),
created,
model: model_id,
model: model_id.clone(),
choices: vec![ChatCompletionChunkChoice {
index: 0,
delta: Delta {
role: None,
content: None,
},
delta: Delta { role: None, content: None },
finish_reason: Some("stop".to_string()),
}],
};
if let Ok(json) = serde_json::to_string(&final_chunk) {
events.push(Ok(Event::default().data(json)));
let _ = tx.send(Ok(Event::default().data(json)));
}
}
// Add [DONE] event
events.push(Ok(Event::default().data("[DONE]")));
// Create a stream from the events
let stream = stream::iter(events);
// Return the SSE stream
let _ = tx.send(Ok(Event::default().data("[DONE]")));
});
// Convert receiver into a Stream for SSE
let stream = UnboundedReceiverStream::new(rx);
Ok(Sse::new(stream))
}
// -------------------------
// Router
// -------------------------
pub fn create_router(app_state: AppState) -> Router {
let cors = CorsLayer::new()
.allow_headers(Any)
.allow_origin(Any)
.allow_methods(Any)
.allow_headers(Any);
Router::new()
.route("/v1/chat/completions", post(chat_completions))
.route("/v1/models", get(list_models))
.layer(cors)
.with_state(app_state)
}
/// Handler for GET /v1/models - returns list of available models
async fn list_models() -> Json<ModelListResponse> {
pub async fn list_models() -> Json<ModelListResponse> {
// Get all available model variants from the Which enum
let models = vec![
Model {
@@ -700,24 +698,6 @@ async fn list_models() -> Json<ModelListResponse> {
})
}
// -------------------------
// Router
// -------------------------
pub fn create_router(app_state: AppState) -> Router {
let cors = CorsLayer::new()
.allow_headers(Any)
.allow_origin(Any)
.allow_methods(Any)
.allow_headers(Any);
Router::new()
.route("/v1/chat/completions", post(chat_completions))
.route("/v1/models", get(list_models))
// .route("/v1/chat/completions/stream", post(chat_completions_stream))
.layer(cors)
.with_state(app_state)
}
#[cfg(test)]
mod tests {
@@ -725,92 +705,59 @@ mod tests {
use crate::openai_types::{Message, MessageContent};
use either::Either;
#[tokio::test]
async fn test_models_list_endpoint() {
println!("[DEBUG_LOG] Testing models list endpoint");
let response = list_models().await;
let models_response = response.0;
// Verify response structure
assert_eq!(models_response.object, "list");
assert_eq!(models_response.data.len(), 16);
// Verify some key models are present
let model_ids: Vec<String> = models_response.data.iter().map(|m| m.id.clone()).collect();
assert!(model_ids.contains(&"gemma-2b".to_string()));
assert!(model_ids.contains(&"gemma-7b".to_string()));
assert!(model_ids.contains(&"gemma-3-1b-it".to_string()));
assert!(model_ids.contains(&"codegemma-2b-it".to_string()));
// Verify model structure
for model in &models_response.data {
assert_eq!(model.object, "model");
assert_eq!(model.owned_by, "google");
assert_eq!(model.created, 1686935002);
assert!(!model.id.is_empty());
}
println!("[DEBUG_LOG] Models list endpoint test passed - {} models available", models_response.data.len());
#[test]
fn test_build_gemma_prompt() {
let messages = vec![
Message {
role: "system".to_string(),
content: Some(MessageContent(Either::Left("System message".to_string()))),
name: None,
},
Message {
role: "user".to_string(),
content: Some(MessageContent(Either::Left("Knock knock.".to_string()))),
name: None,
},
Message {
role: "assistant".to_string(),
content: Some(MessageContent(Either::Left("Who's there?".to_string()))),
name: None,
},
Message {
role: "user".to_string(),
content: Some(MessageContent(Either::Left("Gemma.".to_string()))),
name: None,
},
];
let prompt = build_gemma_prompt(&messages);
let expected = "<start_of_turn>user\nSystem message\n\nKnock knock.<end_of_turn>\n\
<start_of_turn>model\nWho's there?<end_of_turn>\n\
<start_of_turn>user\nGemma.<end_of_turn>\n\
<start_of_turn>model\n";
assert_eq!(prompt, expected);
}
#[tokio::test]
async fn test_reproduce_tensor_shape_mismatch() {
// Create a test app state with Gemma 3 model (same as the failing request)
let mut args = PipelineArgs::default();
args.model_id = "google/gemma-3-1b-it".to_string();
args.which = Which::InstructV3_1B;
println!("[DEBUG_LOG] Creating pipeline with model: {}", args.model_id);
// This should reproduce the same conditions as the curl script
let text_generation = build_pipeline(args.clone());
let app_state = AppState {
text_generation: Arc::new(Mutex::new(text_generation)),
model_id: "gemma-3-1b-it".to_string(),
build_args: args,
};
#[test]
fn test_empty_messages() {
let messages: Vec<Message> = vec![];
let prompt = build_gemma_prompt(&messages);
assert_eq!(prompt, "<start_of_turn>model\n");
}
// Create the same request as the curl script
let request = ChatCompletionRequest {
model: "gemma-3-1b-it".to_string(),
messages: vec![Message {
#[test]
fn test_missing_content() {
let messages = vec![
Message {
role: "user".to_string(),
content: Some(MessageContent(Either::Left("What is the capital of France?".to_string()))),
content: None,
name: None,
}],
max_tokens: Some(128),
stream: Some(true),
temperature: None,
top_p: None,
logprobs: false,
n_choices: 1,
};
}
];
println!("[DEBUG_LOG] Attempting to reproduce tensor shape mismatch error...");
// This should trigger the same error as the curl script
let result = handle_streaming_request(app_state, request).await;
match result {
Ok(_) => {
println!("[DEBUG_LOG] No error occurred - this suggests the issue might be fixed or environmental");
}
Err((status_code, json_error)) => {
println!("[DEBUG_LOG] Error reproduced! Status: {:?}", status_code);
println!("[DEBUG_LOG] Error details: {:?}", json_error);
// Check if this is the expected tensor shape mismatch error
if let Some(error_obj) = json_error.0.as_object() {
if let Some(error_details) = error_obj.get("error").and_then(|e| e.as_object()) {
if let Some(message) = error_details.get("message").and_then(|m| m.as_str()) {
assert!(message.contains("shape mismatch"),
"Expected shape mismatch error, got: {}", message);
println!("[DEBUG_LOG] Successfully reproduced tensor shape mismatch error");
}
}
}
}
}
let prompt = build_gemma_prompt(&messages);
assert_eq!(prompt, "<start_of_turn>user\n<end_of_turn>\n<start_of_turn>model\n");
}
}

295
crates/inference-engine/src/text_generation.rs
View File

@@ -20,6 +20,8 @@ pub struct TextGeneration {
repeat_last_n: usize,
// Cache for repeat penalty computation to avoid redundant calculations
penalty_cache: HashMap<usize, f32>,
// Context window size for sliding window context (default: 64 tokens)
context_window_size: usize,
}
impl TextGeneration {
@@ -55,6 +57,7 @@ impl TextGeneration {
cpu_device,
try_primary_device,
penalty_cache: HashMap::new(),
context_window_size: 64, // Default sliding window size for better context preservation
}
}
@@ -192,9 +195,14 @@ impl TextGeneration {
// Track overall performance
let start_time = std::time::Instant::now();
// Clear penalty cache for new generation
self.penalty_cache.clear();
tracing::debug!("Cleared penalty cache for new generation");
// Keep penalty cache across generation for better repetition prevention
// Only clear cache if it becomes too large to prevent memory bloat
if self.penalty_cache.len() > 10000 {
self.penalty_cache.clear();
tracing::debug!("Cleared penalty cache due to size limit");
} else {
tracing::debug!("Maintaining penalty cache across generation for better repetition prevention");
}
// Phase 1: Tokenize input
let tokenize_start = std::time::Instant::now();
@@ -440,9 +448,14 @@ impl TextGeneration {
// Track overall performance
let start_time = std::time::Instant::now();
// Clear penalty cache for new generation
self.penalty_cache.clear();
tracing::debug!("Cleared penalty cache for new generation (API mode)");
// Keep penalty cache across generation for better repetition prevention
// Only clear cache if it becomes too large to prevent memory bloat
if self.penalty_cache.len() > 10000 {
self.penalty_cache.clear();
tracing::debug!("Cleared penalty cache due to size limit (API mode)");
} else {
tracing::debug!("Maintaining penalty cache across generation for better repetition prevention (API mode)");
}
// Phase 1: Tokenize input
let tokenize_start = std::time::Instant::now();
@@ -573,10 +586,18 @@ impl TextGeneration {
for index in 0..sample_len {
let token_start = std::time::Instant::now();
let context_size = if index > 0 { 1 } else { tokens.len() };
// Use sliding window context instead of single token to preserve context and reduce repetition
let context_size = if index > 0 {
std::cmp::min(self.context_window_size, tokens.len())
} else {
tokens.len()
};
let start_pos = tokens.len().saturating_sub(context_size);
let ctxt = &tokens[start_pos..];
tracing::debug!("API standard model: Using sliding window context: {} tokens (from position {})",
ctxt.len(), start_pos);
// Track tensor operations and model forward pass
let forward_start = std::time::Instant::now();
let input = Tensor::new(ctxt, &self.device)?.unsqueeze(0)?;
@@ -629,6 +650,266 @@ impl TextGeneration {
Ok(())
}
// Run text generation with streaming callback for each token
pub async fn run_with_streaming<F>(&mut self, prompt: &str, sample_len: usize, mut token_callback: F) -> Result<String>
where
F: FnMut(&str) -> Result<()>,
{
// Track overall performance
let start_time = std::time::Instant::now();
// Keep penalty cache across generation for better repetition prevention
// Only clear cache if it becomes too large to prevent memory bloat
if self.penalty_cache.len() > 10000 {
self.penalty_cache.clear();
tracing::debug!("Cleared penalty cache due to size limit (streaming mode)");
} else {
tracing::debug!("Maintaining penalty cache across generation for better repetition prevention (streaming mode)");
}
// Phase 1: Tokenize input
let tokenize_start = std::time::Instant::now();
self.tokenizer.clear();
let mut tokens = self
.tokenizer
.tokenizer()
.encode(prompt, true)
.map_err(E::msg)?
.get_ids()
.to_vec();
let tokenize_time = tokenize_start.elapsed();
tracing::debug!("Streaming Tokenization completed in {:.2?}", tokenize_time);
tracing::debug!("Streaming Input tokens: {}", tokens.len());
// Collect all output for final return
let mut full_output = String::new();
let mut generated_tokens = 0usize;
let eos_token = match self.tokenizer.get_token("<eos>") {
Some(token) => token,
None => anyhow::bail!("cannot find the <eos> token"),
};
let eot_token = match self.tokenizer.get_token("<end_of_turn>") {
Some(token) => token,
None => {
tracing::warn!("Warning: <end_of_turn> token not found in tokenizer, using <eos> as a backup");
eos_token
}
};
// Determine if we're using a Model2 (gemma-2) or Model3 (gemma-3) variant
let needs_special_handling = match &self.model {
Model::V2(_) => true,
Model::V3(_) => true,
_ => false,
};
// Track generation timing
let start_gen = std::time::Instant::now();
// Track per-token generation timing for performance analysis
let mut token_times = Vec::new();
let mut forward_times = Vec::new();
let mut repeat_penalty_times = Vec::new();
let mut sampling_times = Vec::new();
// For Model2 and Model3, we need to use a special approach for shape compatibility
if needs_special_handling {
tracing::debug!("Using special generation approach for gemma-2/gemma-3 models (streaming)");
tracing::debug!("Streaming: sample_len = {}", sample_len);
// Initial generation with the full prompt
let forward_start = std::time::Instant::now();
let input = Tensor::new(tokens.as_slice(), &self.device)?.unsqueeze(0)?;
let mut logits = self.execute_with_fallback(&input, 0)?;
logits = logits.squeeze(0)?.squeeze(0)?.to_dtype(DType::F32)?;
let forward_time = forward_start.elapsed();
forward_times.push(forward_time);
tracing::debug!("Streaming: About to enter generation loop with sample_len = {}", sample_len);
for gen_index in 0..sample_len {
tracing::debug!("Streaming: Starting generation iteration {} / {}", gen_index + 1, sample_len);
let token_start = std::time::Instant::now();
// Apply repeat penalty using optimized cached implementation
let (current_logits, repeat_time) = self.apply_cached_repeat_penalty(logits.clone(), &tokens)?;
repeat_penalty_times.push(repeat_time);
// Track token sampling
let sampling_start = std::time::Instant::now();
let next_token = self.logits_processor.sample(&current_logits)?;
let sampling_time = sampling_start.elapsed();
sampling_times.push(sampling_time);
tokens.push(next_token);
generated_tokens += 1;
tracing::debug!("Streaming: Generated token {} (id: {}), eos: {}, eot: {}",
next_token, next_token, eos_token, eot_token);
if next_token == eos_token || next_token == eot_token {
tracing::debug!("Streaming: Breaking due to end token");
break;
}
if let Some(token_text) = self.tokenizer.next_token(next_token)? {
full_output.push_str(&token_text);
// Call the streaming callback with this token
token_callback(&token_text)?;
}
// For the next iteration, use single token to avoid shape mismatch
let forward_start = std::time::Instant::now();
tracing::debug!("Streaming: Preparing next forward pass with {} tokens", tokens.len());
// Use just the last token for subsequent iterations to avoid shape mismatch
// This is required for Gemma model's attention mechanism compatibility
let context_tokens = &tokens[(tokens.len()-1)..];
let start_pos = tokens.len() - 1;
tracing::debug!("Streaming: Using single token context for Gemma: {} tokens (from position {})",
context_tokens.len(), start_pos);
let new_input = match Tensor::new(context_tokens, &self.device) {
Ok(tensor) => tensor,
Err(e) => {
tracing::error!("Streaming: Failed to create input tensor: {}", e);
return Err(e.into());
}
};
let new_input = match new_input.unsqueeze(0) {
Ok(tensor) => tensor,
Err(e) => {
tracing::error!("Streaming: Failed to unsqueeze input tensor: {}", e);
return Err(e.into());
}
};
tracing::debug!("Streaming: About to call execute_with_fallback for iteration {} with start_pos {}", gen_index + 1, start_pos);
logits = match self.execute_with_fallback(&new_input, start_pos) {
Ok(result) => result,
Err(e) => {
tracing::error!("Streaming: Forward pass failed: {}", e);
return Err(e);
}
};
logits = match logits.squeeze(0) {
Ok(result) => result,
Err(e) => {
tracing::error!("Streaming: Failed to squeeze logits (dim 0): {}", e);
return Err(e.into());
}
};
logits = match logits.squeeze(0) {
Ok(result) => result,
Err(e) => {
tracing::error!("Streaming: Failed to squeeze logits (dim 0 again): {}", e);
return Err(e.into());
}
};
logits = match logits.to_dtype(DType::F32) {
Ok(result) => result,
Err(e) => {
tracing::error!("Streaming: Failed to convert logits to F32: {}", e);
return Err(e.into());
}
};
let forward_time = forward_start.elapsed();
forward_times.push(forward_time);
tracing::debug!("Streaming: Forward pass completed for iteration {}", gen_index + 1);
let token_time = token_start.elapsed();
token_times.push(token_time);
// Yield to allow other async tasks to run
tokio::task::yield_now().await;
}
} else {
// Standard approach for other models
tracing::debug!("Using standard generation approach (streaming)");
for index in 0..sample_len {
let token_start = std::time::Instant::now();
// Use sliding window context instead of single token to preserve context and reduce repetition
let context_size = if index > 0 {
std::cmp::min(self.context_window_size, tokens.len())
} else {
tokens.len()
};
let start_pos = tokens.len().saturating_sub(context_size);
let ctxt = &tokens[start_pos..];
tracing::debug!("Standard model: Using sliding window context: {} tokens (from position {})",
ctxt.len(), start_pos);
// Track tensor operations and model forward pass
let forward_start = std::time::Instant::now();
let input = Tensor::new(ctxt, &self.device)?.unsqueeze(0)?;
let logits = self.execute_with_fallback(&input, start_pos)?;
let logits = logits.squeeze(0)?.squeeze(0)?.to_dtype(DType::F32)?;
let forward_time = forward_start.elapsed();
forward_times.push(forward_time);
// Apply repeat penalty using optimized cached implementation
let (logits, repeat_time) = self.apply_cached_repeat_penalty(logits, &tokens)?;
repeat_penalty_times.push(repeat_time);
// Track token sampling
let sampling_start = std::time::Instant::now();
let next_token = self.logits_processor.sample(&logits)?;
let sampling_time = sampling_start.elapsed();
sampling_times.push(sampling_time);
tokens.push(next_token);
generated_tokens += 1;
if next_token == eos_token || next_token == eot_token {
break;
}
if let Some(token_text) = self.tokenizer.next_token(next_token)? {
full_output.push_str(&token_text);
// Call the streaming callback with this token
token_callback(&token_text)?;
}
let token_time = token_start.elapsed();
token_times.push(token_time);
}
}
let dt = start_gen.elapsed();
// Phase 3: Final decoding
let decode_start = std::time::Instant::now();
// Decode any remaining tokens but don't send through callback to avoid repetition
// The tokens were already streamed individually in the generation loop above
if let Some(rest) = self.tokenizer.decode_rest().map_err(E::msg)? {
full_output.push_str(&rest);
// Note: NOT calling token_callback(&rest) here to prevent token repetition
// Individual tokens were already streamed via the callback in the generation loop
}
let decode_time = decode_start.elapsed();
// Log performance metrics
Self::log_performance_metrics(
dt, generated_tokens, &token_times, &forward_times,
&repeat_penalty_times, &sampling_times, tokenize_time,
decode_time, start_time, "Streaming"
);
Ok(full_output)
}
// Helper function for logging performance metrics
fn log_performance_metrics(

3
crates/inference-engine/src/token_output_stream.rs
View File

@@ -40,7 +40,8 @@ impl TokenOutputStream {
};
self.tokens.push(token);
let text = self.decode(&self.tokens[self.prev_index..])?;
if text.len() > prev_text.len() && text.chars().last().unwrap().is_alphanumeric() {
if text.len() > prev_text.len() {
// Modified to include all tokens, not just alphanumeric ones
let text = text.split_at(prev_text.len());
self.prev_index = self.current_index;
self.current_index = self.tokens.len();

11
crates/leptos-chat/Cargo.toml
View File

@@ -49,4 +49,13 @@ features = [
"fast-rng", # Use a faster (but still sufficiently random) RNG
"macro-diagnostics", # Enable better diagnostics for compile-time UUIDs
"js", # Enable JavaScript RNG for WASM targets
]
]
[package.metadata.kube]
image = "ghcr.io/geoffsee/leptos-chat:latest"
replicas = 1
port = 8788
resources.cpu = "500m"
resources.memory = "256Mi"
#ingress.host = "my-service.example.com"
#env = { RUST_LOG = "info", DATABASE_URL = "postgres://..." }

149
crates/leptos-chat/src/lib.rs
View File

@@ -15,7 +15,7 @@ use async_openai_wasm::{
Client,
};
use async_openai_wasm::config::OpenAIConfig;
use async_openai_wasm::types::{ChatCompletionResponseStream, Model};
use async_openai_wasm::types::{ChatCompletionResponseStream, Model, Role, FinishReason};
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct Message {
@@ -127,7 +127,7 @@ async fn fetch_available_models() -> Result<Vec<OpenAIModel>, String> {
}
async fn send_chat_request(chat_request: ChatRequest) -> ChatCompletionResponseStream {
let config = OpenAIConfig::new().with_api_base("http://localhost:8080".to_string());
let config = OpenAIConfig::new().with_api_base("http://localhost:8080/v1".to_string());
let client = Client::with_config(config);
let mut typed_chat = async_openai_wasm::types::CreateChatCompletionRequest {
@@ -205,7 +205,7 @@ async fn send_chat_request(chat_request: ChatRequest) -> ChatCompletionResponseS
// Err("leptos-chat chat request only supported on wasm32 target".to_string())
// }
const DEFAULT_MODEL: &str = "gemma-2b-it";
const DEFAULT_MODEL: &str = "default";
#[component]
fn ChatInterface() -> impl IntoView {
@@ -272,11 +272,37 @@ fn ChatInterface() -> impl IntoView {
let history_count = messages.with_untracked(|msgs| {
let count = msgs.len();
for msg in msgs.iter() {
let message = ChatCompletionRequestUserMessageArgs::default()
.content(msg.content.clone())
.build()
.expect("failed to build message");
chat_messages.push(message.into());
match msg.role.as_str() {
"user" => {
let message = ChatCompletionRequestUserMessageArgs::default()
.content(msg.content.clone())
.build()
.expect("failed to build user message");
chat_messages.push(message.into());
}
"assistant" => {
let message = ChatCompletionRequestAssistantMessageArgs::default()
.content(msg.content.clone())
.build()
.expect("failed to build assistant message");
chat_messages.push(message.into());
}
"system" => {
let message = ChatCompletionRequestSystemMessageArgs::default()
.content(msg.content.clone())
.build()
.expect("failed to build system message");
chat_messages.push(message.into());
}
_ => {
// Default to user message for unknown roles
let message = ChatCompletionRequestUserMessageArgs::default()
.content(msg.content.clone())
.build()
.expect("failed to build default message");
chat_messages.push(message.into());
}
}
}
count
});
@@ -319,51 +345,69 @@ fn ChatInterface() -> impl IntoView {
Ok(mut stream) => {
log::info!("[DEBUG_LOG] send_message: Successfully created stream, starting to receive response");
// Insert a placeholder assistant message to append into
let assistant_id = Uuid::new_v4().to_string();
set_messages.update(|msgs| {
msgs.push_back(Message {
id: assistant_id.clone(),
role: "assistant".to_string(),
content: String::new(),
timestamp: Date::now(),
});
});
// Defer creating assistant message until we receive role=assistant from the stream
let mut assistant_created = false;
let mut content_appended = false;
let mut chunks_received = 0;
// Stream loop: append deltas to the last message
// Stream loop: handle deltas and finish events
while let Some(next) = stream.next().await {
match next {
Ok(chunk) => {
chunks_received += 1;
// Try to pull out the content delta in a tolerant way.
// async-openai 0.28.x stream chunk usually looks like:
// choices[0].delta.content: Option<String>
let mut delta_txt = String::new();
if let Some(choice) = chunk.choices.get(0) {
// Newer message API may expose different shapes; try common ones
// 1) Simple string content delta
if let Some(content) = &choice.delta.content {
delta_txt.push_str(content);
}
// 2) Some providers pack text under .delta.role/.delta.<other>
// If nothing extracted, ignore quietly.
// If a finish_reason arrives, we could stop early,
// but usually the stream naturally ends.
}
if !delta_txt.is_empty() {
set_messages.update(|msgs| {
if let Some(last) = msgs.back_mut() {
if last.role == "assistant" {
last.content.push_str(&delta_txt);
last.timestamp = Date::now();
// 1) Create assistant message when role arrives
if !assistant_created {
if let Some(role) = &choice.delta.role {
if role == &Role::Assistant {
assistant_created = true;
let assistant_id = Uuid::new_v4().to_string();
set_messages.update(|msgs| {
msgs.push_back(Message {
id: assistant_id,
role: "assistant".to_string(),
content: String::new(),
timestamp: Date::now(),
});
});
}
}
});
}
// 2) Append content tokens when provided
if let Some(content) = &choice.delta.content {
if !content.is_empty() {
// If content arrives before role, create assistant message now
if !assistant_created {
assistant_created = true;
let assistant_id = Uuid::new_v4().to_string();
set_messages.update(|msgs| {
msgs.push_back(Message {
id: assistant_id,
role: "assistant".to_string(),
content: String::new(),
timestamp: Date::now(),
});
});
}
content_appended = true;
set_messages.update(|msgs| {
if let Some(last) = msgs.back_mut() {
if last.role == "assistant" {
last.content.push_str(content);
last.timestamp = Date::now();
}
}
});
}
}
// 3) Stop on finish_reason=="stop" (mirrors [DONE])
if let Some(reason) = &choice.finish_reason {
if reason == &FinishReason::Stop {
log::info!("[DEBUG_LOG] send_message: Received finish_reason=stop after {} chunks", chunks_received);
break;
}
}
}
}
Err(e) => {
@@ -381,6 +425,21 @@ fn ChatInterface() -> impl IntoView {
}
}
}
// Cleanup: If we created an assistant message but no content ever arrived, remove the empty message
if assistant_created && !content_appended {
set_messages.update(|msgs| {
let should_pop = msgs
.back()
.map(|m| m.role == "assistant" && m.content.is_empty())
.unwrap_or(false);
if should_pop {
log::info!("[DEBUG_LOG] send_message: Removing empty assistant message (no content received)");
msgs.pop_back();
}
});
}
log::info!("[DEBUG_LOG] send_message: Stream completed successfully, received {} chunks", chunks_received);
}
Err(e) => {

10
crates/predict-otron-9000/Cargo.toml
View File

@@ -24,3 +24,13 @@ embeddings-engine = { path = "../embeddings-engine" }
# Dependencies for inference functionality
inference-engine = { path = "../inference-engine" }
[package.metadata.kube]
image = "ghcr.io/geoffsee/predict-otron-9000:latest"
replicas = 1
port = 8080
resources.cpu = "500m"
resources.memory = "256Mi"
#ingress.host = "my-service.example.com"
#env = { RUST_LOG = "info", DATABASE_URL = "postgres://..." }

33
package-lock.json generated Normal file
View File

@@ -0,0 +1,33 @@
{
"name": "predict-otron-9000",
"lockfileVersion": 3,
"requires": true,
"packages": {
"": {
"dependencies": {
"openai": "^5.16.0"
}
},
"node_modules/openai": {
"version": "5.16.0",
"resolved": "https://registry.npmjs.org/openai/-/openai-5.16.0.tgz",
"integrity": "sha512-hoEH8ZNvg1HXjU9mp88L/ZH8O082Z8r6FHCXGiWAzVRrEv443aI57qhch4snu07yQydj+AUAWLenAiBXhu89Tw==",
"license": "Apache-2.0",
"bin": {
"openai": "bin/cli"
},
"peerDependencies": {
"ws": "^8.18.0",
"zod": "^3.23.8"
},
"peerDependenciesMeta": {
"ws": {
"optional": true
},
"zod": {
"optional": true
}
}
}
}
}

5
package.json Normal file
View File

@@ -0,0 +1,5 @@
{
"dependencies": {
"openai": "^5.16.0"
}
}

48
server.log Normal file
View File

@@ -0,0 +1,48 @@
Compiling inference-engine v0.1.0 (/Users/williamseemueller/workspace/seemueller-io/predict-otron-9000/crates/inference-engine)
warning: unused import: `Config as Config1`
--> crates/inference-engine/src/model.rs:2:42
|
2 | use candle_transformers::models::gemma::{Config as Config1, Model as Model1};
| ^^^^^^^^^^^^^^^^^
|
= note: `#[warn(unused_imports)]` on by default
warning: unused import: `Config as Config2`
--> crates/inference-engine/src/model.rs:3:43
|
3 | use candle_transformers::models::gemma2::{Config as Config2, Model as Model2};
| ^^^^^^^^^^^^^^^^^
warning: unused import: `Config as Config3`
--> crates/inference-engine/src/model.rs:4:43
|
4 | use candle_transformers::models::gemma3::{Config as Config3, Model as Model3};
| ^^^^^^^^^^^^^^^^^
warning: unused import: `self`
--> crates/inference-engine/src/server.rs:10:28
|
10 | use futures_util::stream::{self, Stream};
| ^^^^
warning: `inference-engine` (lib) generated 4 warnings (run `cargo fix --lib -p inference-engine` to apply 4 suggestions)
Compiling predict-otron-9000 v0.1.0 (/Users/williamseemueller/workspace/seemueller-io/predict-otron-9000/crates/predict-otron-9000)
Finished `release` profile [optimized] target(s) in 4.01s
Running `target/release/predict-otron-9000`
2025-08-28T01:43:11.512475Z  INFO predict_otron_9000::middleware::metrics: Performance metrics summary:
avx: false, neon: true, simd128: false, f16c: false
2025-08-28T01:43:11.512811Z  INFO hf_hub: Using token file found "/Users/williamseemueller/.cache/huggingface/token"
retrieved the files in 685.958µs
2025-08-28T01:43:12.661378Z  INFO predict_otron_9000: Unified predict-otron-9000 server listening on 127.0.0.1:8080
2025-08-28T01:43:12.661400Z  INFO predict_otron_9000: Performance metrics tracking enabled - summary logs every 60 seconds
2025-08-28T01:43:12.661403Z  INFO predict_otron_9000: Available endpoints:
2025-08-28T01:43:12.661405Z  INFO predict_otron_9000: GET / - Root endpoint from embeddings-engine
2025-08-28T01:43:12.661407Z  INFO predict_otron_9000: POST /v1/embeddings - Text embeddings
2025-08-28T01:43:12.661409Z  INFO predict_otron_9000: POST /v1/chat/completions - Chat completions
2025-08-28T01:43:19.166677Z  WARN inference_engine::server: Detected repetition pattern: ' plus' (count: 1)
2025-08-28T01:43:19.296257Z  WARN inference_engine::server: Detected repetition pattern: ' plus' (count: 2)
2025-08-28T01:43:19.424883Z  WARN inference_engine::server: Detected repetition pattern: ' plus' (count: 3)
2025-08-28T01:43:19.554508Z  WARN inference_engine::server: Detected repetition pattern: ' plus' (count: 4)
2025-08-28T01:43:19.683153Z  WARN inference_engine::server: Detected repetition pattern: ' plus' (count: 5)
2025-08-28T01:43:19.683181Z  INFO inference_engine::server: Stopping generation due to excessive repetition
2025-08-28T01:43:19.683221Z  INFO inference_engine::server: Text generation stopped: Repetition detected - stopping generation

277
server_fresh.txt Normal file
View File

@@ -0,0 +1,277 @@
warning: unused import: `Config as Config1`
--> crates/inference-engine/src/model.rs:2:42
|
2 | use candle_transformers::models::gemma::{Config as Config1, Model as Model1};
| ^^^^^^^^^^^^^^^^^
|
= note: `#[warn(unused_imports)]` on by default
warning: unused import: `Config as Config2`
--> crates/inference-engine/src/model.rs:3:43
|
3 | use candle_transformers::models::gemma2::{Config as Config2, Model as Model2};
| ^^^^^^^^^^^^^^^^^
warning: unused import: `Config as Config3`
--> crates/inference-engine/src/model.rs:4:43
|
4 | use candle_transformers::models::gemma3::{Config as Config3, Model as Model3};
| ^^^^^^^^^^^^^^^^^
warning: unused import: `self`
--> crates/inference-engine/src/server.rs:10:28
|
10 | use futures_util::stream::{self, Stream};
| ^^^^
warning: `inference-engine` (lib) generated 4 warnings (run `cargo fix --lib -p inference-engine` to apply 4 suggestions)
Finished `release` profile [optimized] target(s) in 0.13s
Running `target/release/predict-otron-9000`
avx: false, neon: true, simd128: false, f16c: false
2025-08-28T00:34:39.293635Z  INFO hf_hub: Using token file found "/Users/williamseemueller/.cache/huggingface/token"
retrieved the files in 295.458µs
2025-08-28T00:34:39.294536Z  INFO predict_otron_9000::middleware::metrics: Performance metrics summary:
2025-08-28T00:34:40.507474Z  INFO predict_otron_9000: Unified predict-otron-9000 server listening on 127.0.0.1:8080
2025-08-28T00:34:40.507503Z  INFO predict_otron_9000: Performance metrics tracking enabled - summary logs every 60 seconds
2025-08-28T00:34:40.507508Z  INFO predict_otron_9000: Available endpoints:
2025-08-28T00:34:40.507512Z  INFO predict_otron_9000: GET / - Root endpoint from embeddings-engine
2025-08-28T00:34:40.507515Z  INFO predict_otron_9000: POST /v1/embeddings - Text embeddings
2025-08-28T00:34:40.507517Z  INFO predict_otron_9000: POST /v1/chat/completions - Chat completions
2025-08-28T00:34:52.313606Z DEBUG request{method=POST uri=/v1/chat/completions version=HTTP/1.1}: tower_http::trace::on_request: started processing request
2025-08-28T00:34:52.313671Z DEBUG request{method=POST uri=/v1/chat/completions version=HTTP/1.1}: inference_engine::server: Formatted prompt: <start_of_turn>user
You are a helpful assistant who responds thoughtfully and concisely.
Write a paragraph about dogs<end_of_turn>
<start_of_turn>model
2025-08-28T00:34:52.313693Z DEBUG request{method=POST uri=/v1/chat/completions version=HTTP/1.1}: predict_otron_9000::middleware::metrics: POST /v1/chat/completions 200 OK - 0 ms
2025-08-28T00:34:52.313709Z DEBUG request{method=POST uri=/v1/chat/completions version=HTTP/1.1}: tower_http::trace::on_response: finished processing request latency=0 ms status=200
2025-08-28T00:34:52.313763Z DEBUG inference_engine::text_generation: Cleared penalty cache for new generation (streaming mode)
2025-08-28T00:34:52.313985Z DEBUG inference_engine::text_generation: Streaming Tokenization completed in 217.04µs
2025-08-28T00:34:52.313990Z DEBUG inference_engine::text_generation: Streaming Input tokens: 26
2025-08-28T00:34:52.340937Z DEBUG inference_engine::text_generation: Using special generation approach for gemma-2/gemma-3 models (streaming)
2025-08-28T00:34:52.602691Z DEBUG inference_engine::server: Streaming token: 'Dogs'
2025-08-28T00:34:52.602718Z DEBUG inference_engine::server: Sending chunk with content: 'Dogs'
2025-08-28T00:34:52.769918Z DEBUG inference_engine::server: Streaming token: ' have'
2025-08-28T00:34:52.769949Z DEBUG inference_engine::server: Sending chunk with content: ' have'
2025-08-28T00:34:52.905947Z DEBUG inference_engine::server: Streaming token: ' captivated'
2025-08-28T00:34:52.905977Z DEBUG inference_engine::server: Sending chunk with content: ' captivated'
2025-08-28T00:34:53.040888Z DEBUG inference_engine::server: Streaming token: ' humans'
2025-08-28T00:34:53.040921Z DEBUG inference_engine::server: Sending chunk with content: ' humans'
2025-08-28T00:34:53.177116Z DEBUG inference_engine::server: Streaming token: ' for'
2025-08-28T00:34:53.177145Z DEBUG inference_engine::server: Sending chunk with content: ' for'
2025-08-28T00:34:53.313887Z DEBUG inference_engine::server: Streaming token: ' millennia'
2025-08-28T00:34:53.313920Z DEBUG inference_engine::server: Sending chunk with content: ' millennia'
2025-08-28T00:34:53.444031Z DEBUG inference_engine::server: Streaming token: ','
2025-08-28T00:34:53.444060Z DEBUG inference_engine::server: Sending chunk with content: ','
2025-08-28T00:34:53.571919Z DEBUG inference_engine::server: Streaming token: ' evolving'
2025-08-28T00:34:53.571951Z DEBUG inference_engine::server: Sending chunk with content: ' evolving'
2025-08-28T00:34:53.699811Z DEBUG inference_engine::server: Streaming token: ' from'
2025-08-28T00:34:53.699852Z DEBUG inference_engine::server: Sending chunk with content: ' from'
2025-08-28T00:34:53.828082Z DEBUG inference_engine::server: Streaming token: ' wolves'
2025-08-28T00:34:53.828111Z DEBUG inference_engine::server: Sending chunk with content: ' wolves'
2025-08-28T00:34:53.957276Z DEBUG inference_engine::server: Streaming token: ' to'
2025-08-28T00:34:53.957313Z DEBUG inference_engine::server: Sending chunk with content: ' to'
2025-08-28T00:34:54.093248Z DEBUG inference_engine::server: Streaming token: ' beloved'
2025-08-28T00:34:54.093284Z DEBUG inference_engine::server: Sending chunk with content: ' beloved'
2025-08-28T00:34:54.228357Z DEBUG inference_engine::server: Streaming token: ' companions'
2025-08-28T00:34:54.228385Z DEBUG inference_engine::server: Sending chunk with content: ' companions'
2025-08-28T00:34:54.356315Z DEBUG inference_engine::server: Streaming token: ' offering'
2025-08-28T00:34:54.356349Z DEBUG inference_engine::server: Sending chunk with content: ' offering'
2025-08-28T00:34:54.484051Z DEBUG inference_engine::server: Streaming token: ' unwavering'
2025-08-28T00:34:54.484085Z DEBUG inference_engine::server: Sending chunk with content: ' unwavering'
2025-08-28T00:34:54.613022Z DEBUG inference_engine::server: Streaming token: ' loyalty'
2025-08-28T00:34:54.613061Z DEBUG inference_engine::server: Sending chunk with content: ' loyalty'
2025-08-28T00:34:54.742024Z DEBUG inference_engine::server: Streaming token: ' alongside'
2025-08-28T00:34:54.742043Z DEBUG inference_engine::server: Sending chunk with content: ' alongside'
2025-08-28T00:34:54.869804Z DEBUG inference_engine::server: Streaming token: ' boundless'
2025-08-28T00:34:54.869829Z DEBUG inference_engine::server: Sending chunk with content: ' boundless'
2025-08-28T00:34:54.998140Z DEBUG inference_engine::server: Streaming token: ' affection'
2025-08-28T00:34:54.998165Z DEBUG inference_engine::server: Sending chunk with content: ' affection'
2025-08-28T00:34:55.126560Z DEBUG inference_engine::server: Streaming token: ' –'
2025-08-28T00:34:55.126582Z DEBUG inference_engine::server: Sending chunk with content: ' –'
2025-08-28T00:34:55.255214Z DEBUG inference_engine::server: Streaming token: ' often'
2025-08-28T00:34:55.255232Z DEBUG inference_engine::server: Sending chunk with content: ' often'
2025-08-28T00:34:55.383529Z DEBUG inference_engine::server: Streaming token: ' fueled'
2025-08-28T00:34:55.383551Z DEBUG inference_engine::server: Sending chunk with content: ' fueled'
2025-08-28T00:34:55.511437Z DEBUG inference_engine::server: Streaming token: ' by'
2025-08-28T00:34:55.511456Z DEBUG inference_engine::server: Sending chunk with content: ' by'
2025-08-28T00:34:55.639748Z DEBUG inference_engine::server: Streaming token: ' their'
2025-08-28T00:34:55.639768Z DEBUG inference_engine::server: Sending chunk with content: ' their'
2025-08-28T00:34:55.767723Z DEBUG inference_engine::server: Streaming token: ' incredible'
2025-08-28T00:34:55.767741Z DEBUG inference_engine::server: Sending chunk with content: ' incredible'
2025-08-28T00:34:55.895796Z DEBUG inference_engine::server: Streaming token: ' ability'
2025-08-28T00:34:55.895817Z DEBUG inference_engine::server: Sending chunk with content: ' ability'
2025-08-28T00:34:56.025191Z DEBUG inference_engine::server: Streaming token: ' at'
2025-08-28T00:34:56.025219Z DEBUG inference_engine::server: Sending chunk with content: ' at'
2025-08-28T00:34:56.153604Z DEBUG inference_engine::server: Streaming token: ' understanding'
2025-08-28T00:34:56.153626Z DEBUG inference_engine::server: Sending chunk with content: ' understanding'
2025-08-28T00:34:56.282571Z DEBUG inference_engine::server: Streaming token: ' human'
2025-08-28T00:34:56.282590Z DEBUG inference_engine::server: Sending chunk with content: ' human'
2025-08-28T00:34:56.411224Z DEBUG inference_engine::server: Streaming token: ' emotion'
2025-08-28T00:34:56.411247Z DEBUG inference_engine::server: Sending chunk with content: ' emotion'
2025-08-28T00:34:56.540028Z DEBUG inference_engine::server: Streaming token: ' through'
2025-08-28T00:34:56.540050Z DEBUG inference_engine::server: Sending chunk with content: ' through'
2025-08-28T00:34:56.668612Z DEBUG inference_engine::server: Streaming token: ' subtle'
2025-08-28T00:34:56.668630Z DEBUG inference_engine::server: Sending chunk with content: ' subtle'
2025-08-28T00:34:56.797698Z DEBUG inference_engine::server: Streaming token: ' cues'
2025-08-28T00:34:56.797716Z DEBUG inference_engine::server: Sending chunk with content: ' cues'
2025-08-28T00:34:56.927032Z DEBUG inference_engine::server: Streaming token: '!'
2025-08-28T00:34:56.927054Z DEBUG inference_engine::server: Sending chunk with content: '!'
2025-08-28T00:34:57.054903Z DEBUG inference_engine::server: Streaming token: ' Beyond'
2025-08-28T00:34:57.054922Z DEBUG inference_engine::server: Sending chunk with content: ' Beyond'
2025-08-28T00:34:57.183890Z DEBUG inference_engine::server: Streaming token: ' companionship'
2025-08-28T00:34:57.183914Z DEBUG inference_engine::server: Sending chunk with content: ' companionship'
2025-08-28T00:34:57.313258Z DEBUG inference_engine::server: Streaming token: ' they'
2025-08-28T00:34:57.313278Z DEBUG inference_engine::server: Sending chunk with content: ' they'
2025-08-28T00:34:57.441875Z DEBUG inference_engine::server: Streaming token: ' provide'
2025-08-28T00:34:57.441897Z DEBUG inference_engine::server: Sending chunk with content: ' provide'
2025-08-28T00:34:57.569839Z DEBUG inference_engine::server: Streaming token: ' crucial'
2025-08-28T00:34:57.569864Z DEBUG inference_engine::server: Sending chunk with content: ' crucial'
2025-08-28T00:34:57.700161Z DEBUG inference_engine::server: Streaming token: ' assistance'
2025-08-28T00:34:57.700184Z DEBUG inference_engine::server: Sending chunk with content: ' assistance'
2025-08-28T00:34:57.828427Z DEBUG inference_engine::server: Streaming token: ' with'
2025-08-28T00:34:57.828453Z DEBUG inference_engine::server: Sending chunk with content: ' with'
2025-08-28T00:34:57.957703Z DEBUG inference_engine::server: Streaming token: ' tasks'
2025-08-28T00:34:57.957727Z DEBUG inference_engine::server: Sending chunk with content: ' tasks'
2025-08-28T00:34:58.085556Z DEBUG inference_engine::server: Streaming token: ' like'
2025-08-28T00:34:58.085579Z DEBUG inference_engine::server: Sending chunk with content: ' like'
2025-08-28T00:34:58.213727Z DEBUG inference_engine::server: Streaming token: ' guarding'
2025-08-28T00:34:58.213750Z DEBUG inference_engine::server: Sending chunk with content: ' guarding'
2025-08-28T00:34:58.342674Z DEBUG inference_engine::server: Streaming token: ' property'
2025-08-28T00:34:58.342696Z DEBUG inference_engine::server: Sending chunk with content: ' property'
2025-08-28T00:34:58.474992Z DEBUG inference_engine::server: Streaming token: ' or'
2025-08-28T00:34:58.475011Z DEBUG inference_engine::server: Sending chunk with content: ' or'
2025-08-28T00:34:58.603613Z DEBUG inference_engine::server: Streaming token: ' assisting'
2025-08-28T00:34:58.603636Z DEBUG inference_engine::server: Sending chunk with content: ' assisting'
2025-08-28T00:34:58.732292Z DEBUG inference_engine::server: Streaming token: ' individuals'
2025-08-28T00:34:58.732316Z DEBUG inference_engine::server: Sending chunk with content: ' individuals'
2025-08-28T00:34:58.861810Z DEBUG inference_engine::server: Streaming token: ' who'
2025-08-28T00:34:58.861847Z DEBUG inference_engine::server: Sending chunk with content: ' who'
2025-08-28T00:34:58.989748Z DEBUG inference_engine::server: Streaming token: ' are'
2025-08-28T00:34:58.989765Z DEBUG inference_engine::server: Sending chunk with content: ' are'
2025-08-28T00:34:59.118088Z DEBUG inference_engine::server: Streaming token: ' blind'
2025-08-28T00:34:59.118105Z DEBUG inference_engine::server: Sending chunk with content: ' blind'
2025-08-28T00:34:59.246722Z DEBUG inference_engine::server: Streaming token: ' and'
2025-08-28T00:34:59.246746Z DEBUG inference_engine::server: Sending chunk with content: ' and'
2025-08-28T00:34:59.375090Z DEBUG inference_engine::server: Streaming token: ' deaf'
2025-08-28T00:34:59.375119Z DEBUG inference_engine::server: Sending chunk with content: ' deaf'
2025-08-28T00:34:59.503369Z DEBUG inference_engine::server: Streaming token: '.'
2025-08-28T00:34:59.503398Z DEBUG inference_engine::server: Sending chunk with content: '.'
2025-08-28T00:34:59.632352Z DEBUG inference_engine::server: Streaming token: ' Their'
2025-08-28T00:34:59.632374Z DEBUG inference_engine::server: Sending chunk with content: ' Their'
2025-08-28T00:34:59.760656Z DEBUG inference_engine::server: Streaming token: ' diverse'
2025-08-28T00:34:59.760675Z DEBUG inference_engine::server: Sending chunk with content: ' diverse'
2025-08-28T00:34:59.889274Z DEBUG inference_engine::server: Streaming token: ' breeds'
2025-08-28T00:34:59.889293Z DEBUG inference_engine::server: Sending chunk with content: ' breeds'
2025-08-28T00:35:00.018013Z DEBUG inference_engine::server: Streaming token: ' reflect'
2025-08-28T00:35:00.018043Z DEBUG inference_engine::server: Sending chunk with content: ' reflect'
2025-08-28T00:35:00.146874Z DEBUG inference_engine::server: Streaming token: ' a'
2025-08-28T00:35:00.146903Z DEBUG inference_engine::server: Sending chunk with content: ' a'
2025-08-28T00:35:00.275232Z DEBUG inference_engine::server: Streaming token: ' fascinating'
2025-08-28T00:35:00.275257Z DEBUG inference_engine::server: Sending chunk with content: ' fascinating'
2025-08-28T00:35:00.403452Z DEBUG inference_engine::server: Streaming token: ' range'
2025-08-28T00:35:00.403472Z DEBUG inference_engine::server: Sending chunk with content: ' range'
2025-08-28T00:35:00.535110Z DEBUG inference_engine::server: Streaming token: ' of'
2025-08-28T00:35:00.535133Z DEBUG inference_engine::server: Sending chunk with content: ' of'
2025-08-28T00:35:00.663383Z DEBUG inference_engine::server: Streaming token: ' personalities'
2025-08-28T00:35:00.663402Z DEBUG inference_engine::server: Sending chunk with content: ' personalities'
2025-08-28T00:35:00.792808Z DEBUG inference_engine::server: Streaming token: ' shaped'
2025-08-28T00:35:00.792836Z DEBUG inference_engine::server: Sending chunk with content: ' shaped'
2025-08-28T00:35:00.921350Z DEBUG inference_engine::server: Streaming token: ' over'
2025-08-28T00:35:00.921378Z DEBUG inference_engine::server: Sending chunk with content: ' over'
2025-08-28T00:35:01.049207Z DEBUG inference_engine::server: Streaming token: ' countless'
2025-08-28T00:35:01.049228Z DEBUG inference_engine::server: Sending chunk with content: ' countless'
2025-08-28T00:35:01.178030Z DEBUG inference_engine::server: Streaming token: ' generations'
2025-08-28T00:35:01.178058Z DEBUG inference_engine::server: Sending chunk with content: ' generations'
2025-08-28T00:35:01.306740Z DEBUG inference_engine::server: Streaming token: '،'
2025-08-28T00:35:01.306762Z DEBUG inference_engine::server: Sending chunk with content: '،'
2025-08-28T00:35:01.434552Z DEBUG inference_engine::server: Streaming token: ' making'
2025-08-28T00:35:01.434573Z DEBUG inference_engine::server: Sending chunk with content: ' making'
2025-08-28T00:35:01.562628Z DEBUG inference_engine::server: Streaming token: ' them'
2025-08-28T00:35:01.562647Z DEBUG inference_engine::server: Sending chunk with content: ' them'
2025-08-28T00:35:01.690509Z DEBUG inference_engine::server: Streaming token: ' truly'
2025-08-28T00:35:01.690530Z DEBUG inference_engine::server: Sending chunk with content: ' truly'
2025-08-28T00:35:01.819330Z DEBUG inference_engine::server: Streaming token: ' unique'
2025-08-28T00:35:01.819351Z DEBUG inference_engine::server: Sending chunk with content: ' unique'
2025-08-28T00:35:01.947700Z DEBUG inference_engine::server: Streaming token: ' members'
2025-08-28T00:35:01.947720Z DEBUG inference_engine::server: Sending chunk with content: ' members'
2025-08-28T00:35:02.076045Z DEBUG inference_engine::server: Streaming token: ' within'
2025-08-28T00:35:02.076071Z DEBUG inference_engine::server: Sending chunk with content: ' within'
2025-08-28T00:35:02.204721Z DEBUG inference_engine::server: Streaming token: ' our'
2025-08-28T00:35:02.204743Z DEBUG inference_engine::server: Sending chunk with content: ' our'
2025-08-28T00:35:02.333483Z DEBUG inference_engine::server: Streaming token: ' families'
2025-08-28T00:35:02.333506Z DEBUG inference_engine::server: Sending chunk with content: ' families'
2025-08-28T00:35:02.461905Z DEBUG inference_engine::server: Streaming token: ','
2025-08-28T00:35:02.461926Z DEBUG inference_engine::server: Sending chunk with content: ','
2025-08-28T00:35:02.589686Z DEBUG inference_engine::server: Streaming token: ' enriching'
2025-08-28T00:35:02.589710Z DEBUG inference_engine::server: Sending chunk with content: ' enriching'
2025-08-28T00:35:02.718589Z DEBUG inference_engine::server: Streaming token: ' lives'
2025-08-28T00:35:02.718618Z DEBUG inference_engine::server: Sending chunk with content: ' lives'
2025-08-28T00:35:02.846614Z DEBUG inference_engine::server: Streaming token: ' in'
2025-08-28T00:35:02.846635Z DEBUG inference_engine::server: Sending chunk with content: ' in'
2025-08-28T00:35:02.976008Z DEBUG inference_engine::server: Streaming token: ' profound'
2025-08-28T00:35:02.976028Z DEBUG inference_engine::server: Sending chunk with content: ' profound'
2025-08-28T00:35:03.107573Z DEBUG inference_engine::server: Streaming token: ' ways'
2025-08-28T00:35:03.107594Z DEBUG inference_engine::server: Sending chunk with content: ' ways'
2025-08-28T00:35:03.236069Z DEBUG inference_engine::server: Streaming token: ' regardless'
2025-08-28T00:35:03.236088Z DEBUG inference_engine::server: Sending chunk with content: ' regardless'
2025-08-28T00:35:03.364469Z DEBUG inference_engine::server: Streaming token: ' if'
2025-08-28T00:35:03.364492Z DEBUG inference_engine::server: Sending chunk with content: ' if'
2025-08-28T00:35:03.492669Z DEBUG inference_engine::server: Streaming token: ' we'
2025-08-28T00:35:03.492690Z DEBUG inference_engine::server: Sending chunk with content: ' we'
2025-08-28T00:35:03.621905Z DEBUG inference_engine::server: Streaming token: ' choose'
2025-08-28T00:35:03.621927Z DEBUG inference_engine::server: Sending chunk with content: ' choose'
2025-08-28T00:35:03.754038Z DEBUG inference_engine::server: Streaming token: ' to'
2025-08-28T00:35:03.754059Z DEBUG inference_engine::server: Sending chunk with content: ' to'
2025-08-28T00:35:03.883044Z DEBUG inference_engine::server: Streaming token: ' own'
2025-08-28T00:35:03.883066Z DEBUG inference_engine::server: Sending chunk with content: ' own'
2025-08-28T00:35:04.010685Z DEBUG inference_engine::server: Streaming token: ' one'
2025-08-28T00:35:04.010703Z DEBUG inference_engine::server: Sending chunk with content: ' one'
2025-08-28T00:35:04.139584Z DEBUG inference_engine::server: Streaming token: ' ourselves'
2025-08-28T00:35:04.139609Z DEBUG inference_engine::server: Sending chunk with content: ' ourselves'
2025-08-28T00:35:04.269128Z DEBUG inference_engine::server: Streaming token: ' truly'
2025-08-28T00:35:04.269144Z DEBUG inference_engine::server: Sending chunk with content: ' truly'
2025-08-28T00:35:04.398132Z DEBUG inference_engine::server: Streaming token: ' truly'
2025-08-28T00:35:04.398151Z DEBUG inference_engine::server: Sending chunk with content: ' truly'
2025-08-28T00:35:04.527627Z DEBUG inference_engine::server: Streaming token: ' truly'
2025-08-28T00:35:04.527654Z DEBUG inference_engine::server: Sending chunk with content: ' truly'
2025-08-28T00:35:04.657885Z DEBUG inference_engine::server: Streaming token: ' truly'
2025-08-28T00:35:04.657914Z DEBUG inference_engine::server: Sending chunk with content: ' truly'
2025-08-28T00:35:04.788586Z DEBUG inference_engine::server: Streaming token: ' truly'
2025-08-28T00:35:04.788607Z DEBUG inference_engine::server: Sending chunk with content: ' truly'
2025-08-28T00:35:04.918153Z DEBUG inference_engine::server: Streaming token: ' truly'
2025-08-28T00:35:04.918179Z DEBUG inference_engine::server: Sending chunk with content: ' truly'
2025-08-28T00:35:05.048431Z DEBUG inference_engine::server: Streaming token: ' truly'
2025-08-28T00:35:05.048460Z DEBUG inference_engine::server: Sending chunk with content: ' truly'
2025-08-28T00:35:05.178022Z DEBUG inference_engine::server: Streaming token: ' truly'
2025-08-28T00:35:05.178055Z DEBUG inference_engine::server: Sending chunk with content: ' truly'
2025-08-28T00:35:05.308805Z DEBUG inference_engine::server: Streaming token: ' truly'
2025-08-28T00:35:05.308833Z DEBUG inference_engine::server: Sending chunk with content: ' truly'
2025-08-28T00:35:05.438091Z DEBUG inference_engine::server: Streaming token: ' truly'
2025-08-28T00:35:05.438113Z DEBUG inference_engine::server: Sending chunk with content: ' truly'
2025-08-28T00:35:05.561745Z  INFO inference_engine::text_generation: Streaming Text generation completed in 13.22s
2025-08-28T00:35:05.561767Z  INFO inference_engine::text_generation: Streaming Tokens generated: 100
2025-08-28T00:35:05.561770Z  INFO inference_engine::text_generation: Streaming Generation speed: 7.56 tokens/second
2025-08-28T00:35:05.561772Z  INFO inference_engine::text_generation: Streaming Average time per token: 129.65ms
2025-08-28T00:35:05.561774Z DEBUG inference_engine::text_generation: Streaming - Forward pass: 124.98ms (96.4%)
2025-08-28T00:35:05.561776Z DEBUG inference_engine::text_generation: Streaming - Repeat penalty: 74.02µs (0.1%)
2025-08-28T00:35:05.561778Z DEBUG inference_engine::text_generation: Streaming - Sampling: 5.85ms (4.5%)
2025-08-28T00:35:05.561779Z  INFO inference_engine::text_generation: Streaming Total request time: 13.25s
2025-08-28T00:35:05.561781Z DEBUG inference_engine::text_generation: Streaming - Tokenization: 217.04µs (0.0%)
2025-08-28T00:35:05.561782Z DEBUG inference_engine::text_generation: Streaming - Generation: 13.22s (99.8%)
2025-08-28T00:35:05.561783Z DEBUG inference_engine::text_generation: Streaming - Final decoding: 8.17µs (0.0%)
2025-08-28T00:35:30.845607Z DEBUG request{method=POST uri=/v1/chat/completions version=HTTP/1.1}: tower_http::trace::on_request: started processing request
2025-08-28T00:35:30.845670Z DEBUG request{method=POST uri=/v1/chat/completions version=HTTP/1.1}: inference_engine::server: Formatted prompt: <start_of_turn>user
You are a helpful assistant who responds thoughtfully and concisely.
Write a paragraph about cats<end_of_turn>
<start_of_turn>model
2025-08-28T00:35:30.845684Z DEBUG request{method=POST uri=/v1/chat/completions version=HTTP/1.1}: predict_otron_9000::middleware::metrics: POST /v1/chat/completions 200 OK - 0 ms
2025-08-28T00:35:30.845691Z DEBUG request{method=POST uri=/v1/chat/completions version=HTTP/1.1}: tower_http::trace::on_response: finished processing request latency=0 ms status=200
2025-08-28T00:35:30.845719Z DEBUG inference_engine::text_generation: Cleared penalty cache for new generation (streaming mode)
2025-08-28T00:35:30.845789Z DEBUG inference_engine::text_generation: Streaming Tokenization completed in 65.50µs
2025-08-28T00:35:30.845794Z DEBUG inference_engine::text_generation: Streaming Input tokens: 26
2025-08-28T00:35:30.871195Z DEBUG inference_engine::text_generation: Using special generation approach for gemma-2/gemma-3 models (streaming)
./run_server.sh: line 7: 30566 Killed: 9 cargo run --bin predict-otron-9000 --release

39
server_log.txt Normal file
View File

@@ -0,0 +1,39 @@
Compiling inference-engine v0.1.0 (/Users/williamseemueller/workspace/seemueller-io/predict-otron-9000/crates/inference-engine)
warning: unused import: `Config as Config1`
--> crates/inference-engine/src/model.rs:2:42
|
2 | use candle_transformers::models::gemma::{Config as Config1, Model as Model1};
| ^^^^^^^^^^^^^^^^^
|
= note: `#[warn(unused_imports)]` on by default
warning: unused import: `Config as Config2`
--> crates/inference-engine/src/model.rs:3:43
|
3 | use candle_transformers::models::gemma2::{Config as Config2, Model as Model2};
| ^^^^^^^^^^^^^^^^^
warning: unused import: `Config as Config3`
--> crates/inference-engine/src/model.rs:4:43
|
4 | use candle_transformers::models::gemma3::{Config as Config3, Model as Model3};
| ^^^^^^^^^^^^^^^^^
warning: unused import: `self`
--> crates/inference-engine/src/server.rs:10:28
|
10 | use futures_util::stream::{self, Stream};
| ^^^^
warning: `inference-engine` (lib) generated 4 warnings (run `cargo fix --lib -p inference-engine` to apply 4 suggestions)
Compiling predict-otron-9000 v0.1.0 (/Users/williamseemueller/workspace/seemueller-io/predict-otron-9000/crates/predict-otron-9000)
Finished `release` profile [optimized] target(s) in 4.24s
Running `target/release/predict-otron-9000`
avx: false, neon: true, simd128: false, f16c: false
2025-08-28T00:28:26.075133Z  INFO hf_hub: Using token file found "/Users/williamseemueller/.cache/huggingface/token"
retrieved the files in 557.625µs
2025-08-28T00:28:26.075815Z  INFO predict_otron_9000::middleware::metrics: Performance metrics summary:
thread 'main' panicked at crates/predict-otron-9000/src/main.rs:91:61:
called `Result::unwrap()` on an `Err` value: Os { code: 48, kind: AddrInUse, message: "Address already in use" }
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

69
test_predict_otron.sh Executable file
View File

@@ -0,0 +1,69 @@
#!/bin/bash
# Script to test predict-otron-9000 server with 2 sequential CLI requests
# Ensures proper cleanup of child processes on exit
set -e # Exit on any error
# Function to cleanup background processes
cleanup() {
echo "[INFO] Cleaning up background processes..."
if [[ -n "$SERVER_PID" ]]; then
echo "[INFO] Killing server process (PID: $SERVER_PID)"
kill "$SERVER_PID" 2>/dev/null || true
wait "$SERVER_PID" 2>/dev/null || true
fi
# Kill any remaining cargo processes related to predict-otron-9000
pkill -f "predict-otron-9000" 2>/dev/null || true
echo "[INFO] Cleanup complete"
}
# Set up trap to ensure cleanup on script exit
trap cleanup EXIT INT TERM
# Set environment variables
export SERVER_PORT=${SERVER_PORT:-8080}
export RUST_LOG=${RUST_LOG:-info}
echo "[INFO] Starting predict-otron-9000 server in background..."
# Start the server in background and capture its PID
cargo run --bin predict-otron-9000 --release > server.log 2>&1 &
SERVER_PID=$!
echo "[INFO] Server started with PID: $SERVER_PID"
# Function to check if server is ready
check_server() {
curl -s -f http://localhost:8080/v1/models > /dev/null 2>&1
}
# Wait for server to be ready
echo "[INFO] Waiting for server to be ready..."
TIMEOUT=60 # 60 seconds timeout
ELAPSED=0
while ! check_server; do
if [[ $ELAPSED -ge $TIMEOUT ]]; then
echo "[ERROR] Server did not start within $TIMEOUT seconds"
exit 1
fi
sleep 2
ELAPSED=$((ELAPSED + 2))
echo "[INFO] Still waiting for server... (${ELAPSED}s elapsed)"
done
echo "[INFO] Server is ready!"
# Run first CLI request
echo "[INFO] Running first CLI request - listing models..."
bun run cli.ts --list-models
echo ""
echo "[INFO] Running second CLI request - chat completion..."
bun run cli.ts "What is 2+2?"
echo ""
echo "[INFO] Both CLI requests completed successfully!"

85
test_repetition.ts Normal file
View File

@@ -0,0 +1,85 @@
#!/usr/bin/env node
// Test script to reproduce token repetition issue with special characters
const { fetch } = require('node-fetch');
async function testTokenRepetition() {
console.log("Testing token repetition with special characters...");
try {
const response = await fetch('http://localhost:8080/chat/stream', {
method: 'POST',
headers: {
'Content-Type': 'application/json',
},
body: JSON.stringify({
message: "Write a simple greeting with punctuation marks like: Hello! How are you? I'm fine, thanks."
})
});
if (!response.ok) {
throw new Error(`HTTP error! status: ${response.status}`);
}
const reader = response.body?.getReader();
if (!reader) {
throw new Error('No reader available');
}
let fullResponse = '';
let tokens = [];
while (true) {
const { done, value } = await reader.read();
if (done) break;
const chunk = new TextDecoder().decode(value);
const lines = chunk.split('\n');
for (const line of lines) {
if (line.startsWith('data: ')) {
const data = line.slice(6);
if (data === '[DONE]') {
continue;
}
try {
const parsed = JSON.parse(data);
if (parsed.token) {
tokens.push(parsed.token);
fullResponse += parsed.token;
console.log(`Token: "${parsed.token}"`);
}
} catch (e) {
console.log(`Non-JSON data: ${data}`);
}
}
}
}
console.log('\n=== ANALYSIS ===');
console.log('Full response:', fullResponse);
console.log('Total tokens:', tokens.length);
// Check for repetition issues
const tokenCounts = {};
let hasRepetition = false;
for (const token of tokens) {
tokenCounts[token] = (tokenCounts[token] || 0) + 1;
if (tokenCounts[token] > 1 && token.match(/[!?,.;:]/)) {
console.log(`⚠️ Repetition detected: "${token}" appears ${tokenCounts[token]} times`);
hasRepetition = true;
}
}
if (!hasRepetition) {
console.log('✅ No token repetition detected');
}
} catch (error) {
console.error('Error testing token repetition:', error);
}
}
testTokenRepetition();