[Quantization] Add TurboQuant dynamic kv cache compression

tests/kernels/bench_turboquant.py

@@ -0,0 +1,44 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Benchmark TurboQuant vs baseline throughput."""
+
+import time
+
+import torch
+
+from vllm import LLM, SamplingParams
+
+MODEL = "Qwen/Qwen2.5-7B-Instruct"
+PROMPT = "Explain the theory of general relativity in detail."
+
+
+def benchmark(kv_cache_dtype="auto", label="baseline"):
+    print(f"\n{'=' * 50}")
+    print(f"  {label}")
+    print(f"{'=' * 50}")
+    llm = LLM(
+        MODEL,
+        kv_cache_dtype=kv_cache_dtype,
+        max_model_len=4096,
+        gpu_memory_utilization=0.5,
+    )
+    params = SamplingParams(max_tokens=128, temperature=0.0)
+    llm.generate([PROMPT], params)  # warmup
+    torch.accelerator.synchronize()
+    t0 = time.perf_counter()
+    out = llm.generate([PROMPT] * 8, params)
+    torch.accelerator.synchronize()
+    t1 = time.perf_counter()
+    toks = sum(len(o.outputs[0].token_ids) for o in out)
+    tp = toks / (t1 - t0)
+    print(f"  {toks} tokens in {t1 - t0:.2f}s = {tp:.1f} tok/s")
+    print(f"  Sample: {out[0].outputs[0].text[:80]}...")
+    del llm
+    torch.accelerator.empty_cache()
+    return tp
+
+
+if __name__ == "__main__":
+    t1 = benchmark("auto", "BASELINE (bf16)")
+    t2 = benchmark("turboquant", "TURBOQUANT (4-bit)")
+    print(f"\nRatio: {t2 / t1:.2f}x")