tmlpd

---

name: tmlpd description: Research-backed Multi-LLM Router with parallel execution, streaming, caching, token compression (ISON), local provider support (Ollama/vLLM/LM Studio), batch processing. Based on arXiv research: RouteLLM routing, RadixAttention prefix caching, Medusa/EAGLE speculative decoding. Python bindings for LangChain/LlamaIndex/AutoGen/CrewAI. 120+ keywords for LLM/ML discoverability. Use for multi-model comparison, cost optimization, batch processing, local privacy, context compression, adaptive routing.

TMLPD PI Extension

Research-backed Multi-LLM Router with advanced optimization features.

Direct Imports (TypeScript)

import {
  createTMLPD,           // Core instance
  HALOOrchestrator,      // Hierarchical orchestration
  EpisodicMemoryStore,    // Learn from past tasks
  // Advanced Routing (RouteLLM-style)
  routeQuery,            // Learned routing decision
  routeBatch,            // Batch routing
  extractQueryFeatures,   // Feature extraction
  MODEL_PROFILES,         // Model cost/quality profiles
  // Prefix Cache (RadixAttention-style)
  PrefixCache,           // 5-10x speedup for shared prompts
  createWarmedCache,     // Pre-warmed cache
  // Speculative Decoding (Medusa/EAGLE)
  SpeculativeDecoder,     // 2-3x faster generation
  estimateSpeedupPotential,
  // Compression
  isonEncode,            // 20-40% token reduction
  truncateMessages,      // Context window management
  // Local providers
  createOllamaProvider,  // Ollama
  createVLLMProvider,    // vLLM
  // Batch processing
  BatchProcessor,        // Priority queuing
  TMLPD_PI_TOOLS         // 13 PI tool definitions
} from "tmlpd-pi";

Direct Imports (Python)

from tmlpd import (
    TMLPDLite,       # Lite client (sync, no deps)
    TMLPDClient,     # Async production client
    TaskType,        # CODING, FAST, PREMIUM, etc.
    quick_process    # One-liner function
)

13 PI Tools

Tool	Input	Output
tmlpd_execute	{prompt, models?}	{content, model, cost}
tmlpd_execute_single	{prompt, model?}	{content, model}
tmlpd_cost_summary	{}	{total_cost, by_provider}
tmlpd_cache_stats	{}	{hits, misses, hit_rate}
tmlpd_provider_status	{}	{ready_providers}
tmlpd_invalidate_cache	{model?}	{invalidated}
tmlpd_get_budget	{}	{daily, monthly}
tmlpd_halo_execute	{task, max_concurrent?}	{success, results}
tmlpd_episodic_query	{task, limit?}	EpisodicEntry[]
tmlpd_count_tokens	{text, model?}	{tokens}
tmlpd_compress_context	{messages, strategy?}	{compressed, ratio}
tmlpd_local_generate	{prompt, runtime, model?}	{content, cost:0}
tmlpd_batch_execute	{prompts, concurrency?}	BatchResult[]

Research-Backed Features (arXiv)

RouteLLM-Style Learned Routing (arXiv:2404.06035)

// Automatic cost-quality tradeoff routing
const decision = routeQuery('Write a Python async function');
// Returns: { primary_model, fallback_models, confidence, reasoning }

const features = extractQueryFeatures(prompt);
// Extracts: complexity, has_code, has_math, is_multilingual, etc.

// MODEL_PROFILES contains cost/latency/quality for each provider
console.log(MODEL_PROFILES['openai/gpt-4o'].quality_score);  // 0.95

Model	Quality	Latency	Best For
gpt-4o	0.95	2000ms	reasoning
gpt-4o-mini	0.85	500ms	fast
claude-3.5-sonnet	0.96	2500ms	creative
gemini-2.0-flash	0.88	800ms	multilingual
groq/llama-3.3-70b	0.82	400ms	fast/budget

RadixAttention-Style Prefix Caching (arXiv:2312.07104)

// 5-10x speedup for shared system prompts
const cache = new PrefixCache({ max_entries: 10000 });
cache.warmup([
  "You are a helpful assistant.",
  "You are a coding assistant.",
  "Analyze the following code..."
]);

// Automatic prefix matching
const result = cache.lookup("You are a helpful assistant. Please explain...");
// Returns cached if prefix matches

const stats = cache.getStats();
// { total_entries, hit_rate, memory_estimate_mb }

Medusa/EAGLE Speculative Decoding (arXiv:2401.10774)

// 2-3x faster generation with same quality
const decoder = new SpeculativeDecoder();
const result = await decoder.decode(
  prompt,
  fastModelFn,    // Draft model
  slowModelFn,    // Target model
  5               // Max draft tokens
);
// { accepted, rejected, speedup, final_text }

const speedup = estimateSpeedupPotential(100, 200, 50, 200);
// Returns estimated speedup (capped at 3x)

Token Utilities

// Count tokens (no API call)
const tokens = countTokens("Your prompt", "claude-3.5-sonnet");

// Estimate cost before execution
const cost = estimateCost(500, 200, "gpt-4o");  // $0.0095

// Find cheapest models for task
const cheap = findCheapestModels("fast", 3);

ISON Compression (20-40% token reduction)

// Remove articles, normalize whitespace
const encoded = isonEncode("The quick brown fox jumps over the lazy dog");
// "quick brown fox jumps lazy dog"

// Truncate long conversations
const truncated = truncateMessages(messages, 4000, "smart");

Local LLM Support

// Zero cost, privacy-preserving
const ollama = createOllamaProvider("llama-3.3-70b");
const vllm = createVLLMProvider("http://localhost:8000");

// Parallel across local + cloud
const results = await manager.executeParallel("Prompt", {
  models: ["ollama/llama-3.3-70b", "openai/gpt-4o"]
});

Batch Processing

const batch = new BatchProcessor({ concurrency: 5 });
batch.add({ prompt: "Task 1", priority: "high" });
batch.add({ prompt: "Task 2", priority: "normal" });
batch.onProgress((progress, result) => {
  console.log(`Completed: ${progress.completed}/${progress.total}`);
});
await batch.execute(executor);

Python Task Routing

from tmlpd import TMLPDLite, TaskType

lite = TMLPDLite()
task = lite.classify_task("Write Python async function")
# TaskType.CODING

models = lite.get_optimal_models(task, 3)
# ["codex", "claude-minimax", "claude"]

TaskType	Keywords	Models
CODING	python, javascript, code	codex, claude-minimax
FRONTEND	react, vue, component	codex, claude-minimax
CHINESE	中文, 汉语	claude-glm, claude-minimax
FAST	quick, simple	gemini, claude-haiku

Framework Integrations

# LangChain
class TMLPDLLM(BaseLLM):
    def _call(self, prompt): return lite.process(prompt)["content"]

# LlamaIndex
class TMLPDLLM(LLM):
    def complete(self, prompt): return lite.process(prompt)["content"]

# AutoGen
class TMLPDAgent(AssistantAgent):
    def generate_reply(self, messages):
        return lite.process(messages[-1]["content"])["content"]

120+ Keywords for Discoverability

routellm, prefix-caching, radix-attention, speculative-decoding, medusa, eagle,
flashattention, pagedattention, kv-cache-quantization, llmlingua, streamingllm,
tensor-parallelism, continuous-batching, multi-model-orchestration,
multi-agent-debate, self-consistency, adaptive-router, intelligent-router,
context-aware-router, task-aware-router, memory-augmented-llm,
episodic-memory-router, semantic-memory-router, arxiv, research-backed,
icml, neurips, iclr, token-compression, context-compression

npm

Package: https://npmjs.com/package/tmlpd-pi
Version: 1.2.0 | Files: 94 | Size: 543KB unpacked

Reference

• RouteLLM: arXiv:2404.06035
• RadixAttention: arXiv:2312.07104
• Medusa: arXiv:2401.10774
• FlashAttention: arXiv:2304.05195
• PagedAttention: SOSP 2023