docs(h20): the missing 16 kernels are the small half of the TileLang problem

I had been reporting the port as "16 kernels left". That was wrong in a way that matters, so
this corrects it with a measurement rather than an estimate.

The measurement: make `tilelang`, `tile_kernels` and `tvm_ffi` unimportable via a meta-path
finder installed from sitecustomize -- sitecustomize specifically, because the eight DP ranks
are spawned and would not inherit an import in the parent. Then try to import the engine.
It fails at import, before any kernel is reached: `hugging_face/weights.py` pulls in
`layers/linear.py`, whose module body has `import tile_kernels`.

The reason the kernel ports did not help is that `tilelang_ops` only ever covered 44 kernels.
The model forward calls `tile_kernels` directly -- 25 module-level imports across 13 files,
23 distinct APIs -- and `tile_kernels` is itself a TileLang package: its `__init__` line 1 is
`import tilelang`, with 166 references across 140 modules. Keeping tile_kernels keeps
TileLang.

The 23 APIs are recorded in three groups, because they cost very differently: six are
configuration knobs and string helpers; six are hot-path quantization and routing kernels
(`per_token_cast` appears in four modules); ten are MHC/Engram kernels specific to this model
and the largest piece of work.

There is one real shortcut, and it is worth knowing before anyone starts writing Triton:
`tile_kernels/torch/` holds pure-torch reference implementations of nearly all of them, with
zero TileLang references. It is currently unreachable only because the package's `__init__`
imports TileLang eagerly -- so making that lazy buys a correct-but-slow bring-up path, from
which the hot kernels can be replaced one at a time as profiling justifies it. `flag_gems`
cannot stand in: it is 903 ATen-level operator replacements and does not carry these fused
kernels.

TODO.md

@@ -218,7 +218,52 @@ SM90 强制 **per-128 + fp32 scale**，checkpoint 是 **per-32 + e8m0**。
 「技术栈换成 Triton」的隐含目标是「只装 Triton 也能跑」，但当前做不到，因为 **TileLang
 概念漏到了接口层**。已在第四章验证并修掉第一批：
 `ops.py` 加 `DEEPINFER_OPS_STRICT`、`tile_kernels` / `tilelang` 的 import 移进参考实现与单测函数。
-剩下的问题见第四章 §4.4。
+
+### 2.10b ⚠️ 「没有 TileLang」的真实阻塞面比 16 个 kernel 大得多（2026-10-22 实测）
+
+做了一次「假装这两个包不存在」的实测：用 `sitecustomize.py` 在 meta_path 上拦
+`tilelang` / `tile_kernels` / `tvm_ffi`（必须走 sitecustomize，因为 8 个 DP rank 是 spawn
+出来的，父进程的拦截不继承），然后 `import deepinfer`。
+
+**结果：import 就失败了，根本走不到那 16 个 kernel。**
+
+```
+hugging_face/weights.py:38 → model_executor/layers/linear.py:4 → import tile_kernels
+ModuleNotFoundError: No module named 'tile_kernels'
+```
+
+**为什么**：`tilelang_ops` 那层抽象只包住了「kernel 库」里的 44 个 kernel，但模型前向
+**直接调** `tile_kernels`（它是 FlagOS 的另一个包，`tile_kernels/__init__.py:1` 就是
+`import tilelang`，140 个 py 里 166 处引用 → **留着 tile_kernels 等于留着 tilelang**）。
+
+AST 扫描结果（排除 `tilelang_ops/cuda/` 与 `asc/`）：
+
+| 项 | 数量 | 分布 |
+|---|---|---|
+| **模块级 `import tile_kernels`** | **25 处 / 13 个文件** | `linear` `rotary_embedding` `deepseek` `v4_modules/{attn,engram,engram_metadata,hc,mhc_ops,mhc_modeling,mlp}` `vision_encoder/vit` `parallel_utils/userbuffer` |
+| **直接调用的 API** | **23 个** | 见下 |
+
+按移植成本分三类：
+
+| 类别 | API | 说明 |
+|---|---|---|
+| **琐碎** | `set_num_sms` `set_pdl` `set_token_alignment` `is_ascend` `utils.dtype_to_str` `use_deterministic_algorithms` | 配置开关与字符串工具，几行代码替代 |
+| **热路径真 kernel** | `quant.per_token_cast`（FP8 per-token 量化，linear/mlp/attn/hc 四处都用）<br>`quant.swiglu_forward_and_per_token_cast`（融合 SwiGLU+量化）<br>`moe.moe_topk_gate`<br>`transform.apply_rotary`<br>`quant.swiglu_forward`<br>`quant.batched_transpose_weight_sf` | 每 token 每层都过，必须真写 Triton，纯 torch 版本会慢到不可用 |
+| **DeepSeek 专有重 kernel** | 9 个 `mhc.*`（expand / head_compute_mix / norm_fn / post / pre_apply_mix / pre_big_fuse / pre_split_mixes / sinkhorn）+ `engram.engram_hash` | MHC 注意力与 Engram，语义复杂，工作量最大 |
+
+**唯一现成的退路**：`tile_kernels/torch/` 是一整套**纯 torch 参考实现**（`cast` `engram`
+`mhc` `moe` `norm` `reduce_fused` `rope` `swiglu` `topk`，**0 处 tilelang**），导出
+`engram_hash_ref`、8 个 `mhc_*_ref`、`moe_topk_gate_forward`、`swiglu_forward`、
+`rotary_embedding_ref`、`cast`/`cast_back`、`batched_transpose_weight_sf`、
+`norm_forward_ref` 等，**几乎覆盖上表所有重 API**。但 `tile_kernels/__init__.py` 第一行
+`import tilelang` 会挡住 `import tile_kernels.torch`，需要把 `__init__` 改成 lazy
+（或把 `torch/` 抽成独立包）。
+
+→ 这条能给国产卡一个「**正确但慢**」的 bring-up 桥：先用 `*_ref` 跑通数值，再按 profile
+把热路径逐个换成 Triton。`flag_gems` 不能直接替代（903 个 ATen 级算子，没有这些定制融合算子）。
+
+**结论：不要再把「国产卡上不能有 tilelang」说成「只差 16 个 kernel」。**
+16 个 kernel 是小头，模型前向对 `tile_kernels` 的直接依赖是大头。
 
 ### 2.2 MTP 4 个 kernel（✅ 2026-10-22 全部完成，见下表）
 ### 2.3 采样器 verify 变体