test(h20): check what the model actually says, in a file rather than on the command line

`h20_ask.py` takes its prompts as arguments, and a login shell that is not UTF-8 mangles
them into `????` before python ever sees them. The result is a model answering a question
nobody asked, reported as a quality problem. The prompts now live in the script.

Also records, in the plan, the state of the port as measured today: what runs end to end,
the two environment variables it still needs, and the one real blocker found while
checking -- the Engram gather is not CUDA-graph capture safe, so the default
(attention-graph) configuration fails to warm up on eight ranks.
This commit is contained in:
loopyt 2026-10-22 10:53:01 +08:00
parent 615664e772
commit 73b8090c81
2 changed files with 115 additions and 0 deletions

View File

@ -4,6 +4,42 @@
分支:`h20-sm90-triton`,基于 `h20-sm90-support`(PR #180)。
## 零、现状(2026-10-22 09:30 实测)
**能说人话。** 8×H20、`DEEPINFER_OPS_BACKEND=triton`、MTP 默认开(`nextn_predict=3`):
```
=== prompt: '问题:中国的首都是哪里?\n答案:'
=== continuation: '北京。\n问题:中国的首都是哪里?…' # 贪心 64 token 后开始复读提示,正常退化
=== prompt: '请用两句话解释什么是注意力机制。\n答:'
=== continuation: '注意力机制是一种让模型在处理信息时,能够“聚焦”于相关部分而“忽略”无关部分的
技术。它通过计算查询(Query)与键(Key)的相似度来分配权重,再根据权重对值(Value)进行加权求和,
从而得到最终的输出。<|end▁of▁sentence|>…'
```
8 个 DP rank 逐 token 相同,回退提示 5 条(MTP 那 4 个 + 采样器 verify 变体),0 报错。
**但「直接这个状态」= 默认配置跑不起来**,必须两个 env(两个都不是 Triton 引入的):
| env | 为什么必须 |
|---|---|
| `EP_DISABLE_GIN=1` | 容器没挂 `/dev/infiniband`(host 上有 4 个 uverbs),DeepEP 查 `props.ginType` 得 NONE 就 assert。SM90 单机本来走 scale-up,不需要 GIN。 |
| `DEEPINFER_ENFORCE_EAGER=1` | 见下面 P0:CUDA graph capture 里 engram 本地 gather 会 D2H。 |
### 已知问题清单(按优先级)
| # | 问题 | 状态 | 影响 |
|---|---|---|---|
| P0 | **CUDA graph capture 挂掉**:`engram_local.py:108 torch.bincount(owner)`(owner 在 GPU)在 capture 里触发 CPU↔CUDA 拷贝。且 `_exchange_sizes` 用 `all_to_all_single` + 动态 split sizes,本身不可 capture。 | 未修,用 eager 绕过 | 默认配置(graph on)warmup 直接失败,8 rank 全挂 |
| P1 | MTP 4 个 kernel 未移植:`remap_draft_topk` / `get_accept_result` / `finalize_verification` / `get_fused_dflash_get_input` | 未做 | MTP 路径仍有 TileLang 依赖,5 条回退 |
| P1 | 采样器 rejection-sampling 变体(MTP verify)未实现 | 显式委托 cuda | 同上 |
| P2 | 导入期报 `lacks 20 kernel(s)`:这些是**可选路径**(structure_output 引导解码、image_embed_ops 多模态、prepare_cp_mirror CP>1、schedule_mtp_varlen、state_cache、candidate_blocks、sum_smaller_probs、dflash_copy、outlier_detection 等),当前配置不执行,但一旦打开就回退 | 未做 | 功能可用(回退),但「技术栈全 Triton」不成立 |
| P3 | 窄 cut 采样器 token 差 ~9~13%(top_k=128 / top_p=0.5 之类);引擎配置(top_p=0.99)100% | 已定量解释,不修 | 只影响极端参数,分布本身差 1e-5 |
| P3 | `token_logprobs` 恒 0 | 未查 | 与 Triton 无关,历史 TileLang 日志同样为 0;影响 logprob 类接口 |
| P4 | 性能没测 | 未做 | 没有任何 Triton vs TileLang 吞吐数据 |
| — | 2.7% logits 漂移 | SM90 forced per-128 + fp32 scale 的量化地板 | 非 Triton 问题,答案稳健性已验(margin 1.638) |
## 一、范围是怎么定下来的
不靠读代码猜,而是实测。`ops.py` 里加了一个 `DEEPINFER_TRACE_KERNELS=<path>` 探针

View File

@ -0,0 +1,79 @@
"""End-to-end coherence check: a few natural-language prompts, printed as decoded text.
`h20_ask.py` takes prompts on the command line, which a non-UTF-8 login shell mangles into
`????`. The prompts therefore live in this file, so the text the model actually sees is the
text written here.
python scripts/h20_state_check.py --max_tokens 64 # greedy
python scripts/h20_state_check.py --temperature 0.99 # real sampling (sample path)
"""
import argparse
import os
from _config import DEFAULT_MODEL_NAME, create_llm, setup_common_env
from _launch import launch_dp
from deepinfer.sampling_params import SamplingParams
PROMPTS = [
"问题:中国的首都是哪里?\n答案:",
"请用两句话解释什么是注意力机制。\n答:",
"写一句描写秋天的诗。\n",
]
parser = argparse.ArgumentParser()
parser.add_argument("--max_tokens", type=int, default=64)
parser.add_argument("--temperature", type=float, default=0.0)
parser.add_argument("--ep", type=int, default=8)
parser.add_argument("--cp", type=int, default=1)
parser.add_argument("--model_name", default=DEFAULT_MODEL_NAME)
args = parser.parse_args()
ep_size, cp_size = args.ep, args.cp
dp_size = max(ep_size // cp_size, 1)
world_size = dp_size * cp_size
assert world_size % ep_size == 0
def main(dp_size: int, dp_rank: int) -> None:
setup_common_env(
model_name=args.model_name,
sampler_force_topk=4096,
set_override_experts=False,
enable_eplb=False,
)
os.environ.setdefault("DEEPINFER_BATCH_INVARIANT", "1")
llm = create_llm(
args.model_name,
dp_rank,
ep_size=ep_size,
cp_size=cp_size,
env={"DEEPINFER_LOGGING_INTERVAL_STEP": "10"},
data_parallel_size=dp_size,
max_num_batched_tokens=4096,
max_model_len=4096,
max_num_seqs=max(1, len(PROMPTS)),
gpu_memory_utilization=0.85,
lookahead_tokens=0,
)
sampling_params = [
SamplingParams(temperature=args.temperature, max_tokens=args.max_tokens, ignore_eos=True, seq_seed=0) for _ in PROMPTS
]
outputs = llm.generate(prompts=PROMPTS, sampling_params=sampling_params, use_tqdm=False)
for i, output in enumerate(outputs):
assert output, f"sample {i}: no generated tokens"
tokens = [x.token_id for x in output]
text = llm.llm_engine.tokenizer.decode(tokens)
print(f"\n=== DP {dp_rank} sample {i} temp={args.temperature}")
print(f"=== prompt: {PROMPTS[i]!r}")
print(f"=== token ids ({len(tokens)}): {tokens}")
print(f"=== continuation: {text!r}", flush=True)
llm.close()
if __name__ == "__main__":
launch_dp(main, dp_size)