#!/usr/bin/env python3
"""Run the calibration corpus through the live 1M model with the
VLLM_ACT_CAPTURE_DIR hook: per layer, ~24k routed tokens of (x, topk).
Reuses the corpus builder from collect_expert_stats_v2."""
import os
import sys

os.environ["VLLM_ACT_CAPTURE_DIR"] = "/data/glm52-acts"
os.environ.setdefault("VLLM_PP_LAYER_PARTITION", "21,19,19,19")
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))

from collect_expert_stats_v2 import build_corpus  # noqa: E402


def main():
    from vllm import LLM, SamplingParams
    from vllm.inputs import TokensPrompt

    llm = LLM(
        model="/data/glm52",
        pipeline_parallel_size=4,
        gpu_memory_utilization=0.509,
        kv_cache_dtype="fp8_ds_mla",
        max_model_len=8192,
        max_num_seqs=2,
        max_num_batched_tokens=2048,
        enforce_eager=True,
    )
    tok = llm.get_tokenizer()
    texts = build_corpus(tok)
    prompts = [TokensPrompt(prompt_token_ids=tok.encode(t)[:7900])
               for t in texts]
    print(f"{len(prompts)} calibration prompts")
    llm.generate(prompts, SamplingParams(max_tokens=1, temperature=0.0))
    print("generation done; check /data/glm52-acts")


if __name__ == "__main__":
    main()