better

README.md

@@ -4,7 +4,7 @@ A lightweight vLLM implementation built from scratch.
 
 ## Key Features
 
-* 🚀 **Fase offline inference** - Comparable inference speeds to vLLM
+* 🚀 **Fast offline inference** - Comparable inference speeds to vLLM
 * 📖 **Readable codebase** - Clean implementation under 1,200 lines of Python code
 * ⚡ **Optimization Suite** - Prefix caching, Torch compilation, CUDA graph, etc
 

bench.py

@@ -24,4 +24,4 @@ llm.generate(prompt_token_ids, sampling_params)
 t = (time.time() - t)
 total_tokens = sum(sp.max_tokens for sp in sampling_params)
 throughput = total_tokens / t
-print(f"Total: {total_tokens}tok, Time: {t:.2f}s, Throughput: {throughput: .2f}tok/s")
+print(f"Total: {total_tokens}tok, Time: {t:.2f}s, Throughput: {throughput:.2f}tok/s")

nanovllm/engine/block_manager.py

@@ -86,7 +86,7 @@ class BlockManager:
             seq.block_table.append(block_id)
 
     def deallocate(self, seq: Sequence):
-        for block_id in seq.block_table:
+        for block_id in reversed(seq.block_table):
             block = self.blocks[block_id]
             block.ref_count -= 1
             if block.ref_count == 0:

pyproject.toml

@@ -10,7 +10,7 @@ license = "MIT"
 license-files = ["LICENSE"]
 readme = "README.md"
 description = "a mimic VLLM implementation from scratch"
-requires-python = ">=3.9,<3.13"
+requires-python = ">=3.10,<3.13"
 dependencies = [
     "torch>=2.4.0",
     "triton>=3.0.0",