mirror of
https://github.com/deepseek-ai/DeepInfer.git
synced 2026-11-02 17:56:22 +08:00
39b086d781
The references were importable and numerically validated, but nothing pointed the model at them: the forward pass still called the compiled kernels, which raise ModuleNotFoundError on a host without TileLang. This adds the routing layer and swaps the call sites. `deepinfer/tk_compat/` exposes the *compiled* signatures -- out-buffers, in-place conventions, argument order -- because that is what the 25 call sites are written against, and adapts to the reference's shape internally. Selection is `DEEPINFER_KERNELS=auto|cuda|ref`; `auto` follows whether TileLang imports. The adapters are not pass-throughs. Each difference was found by running both backends: * `per_token_cast` takes `num_per_channels=n`, the reference takes `block_size=(1, n)`. * `apply_rotary` takes a concatenated cos|sin cache and rotates in place returning None; the reference takes cos and sin separately and returns a new tensor. The adapter splits, calls, and copies back, because the call site relies on the in-place contract. * `engram_gate` takes the pre-multiplied weight and `(eps, clamp_value)`; the reference needs the two weights separately and `(clamp_value, eps)`. Since a product cannot be un-multiplied, `fused_weight` returns the pair on the reference path and the gate unpacks it. * `mhc_pre_big_fuse_fwd` has no single reference: its five steps are composed from the reference's own components. * `mhc_reduce_partials_and_rmsnorm_fwd` likewise; rebuilt from the reference's formula. `moe_topk_gate(force_random=...)` raises rather than silently ignoring the flag. The reference has no equivalent, and dropping it would route tokens differently while looking healthy. Verification, all on the real call signatures: 18 cases compared between `cuda` and `ref` modes, `ALL_MATCH`, 12 bitwise identical including every quantised byte and every routing index; plus a smoke run that executes all 23 APIs with TileLang blocked, `ALL_CALLS_OK`. End to end still has not completed, and the reason is not this change: six GPUs on the host are held by other containers (82 GiB of 139 each), so the 8-way EP run OOMs in allocate_userbuffer and the surviving ranks wait at a barrier. Recorded in the doc along with the command to rerun it.
26 lines
866 B
Python
26 lines
866 B
Python
"""`tile_kernels.torch` -- the pure-torch reference implementations, unchanged.
|
|
|
|
Some call sites (`engram_metadata`) reach for a helper that only exists on the reference side and
|
|
is pure Python to begin with, e.g. `tile_kernels.torch.engram.make_offsets`. Forwarding the whole
|
|
submodule keeps those working regardless of which backend the kernels resolve to, since there is
|
|
nothing to choose between: the reference *is* the implementation.
|
|
|
|
This also makes the reference reachable for debugging on a host where the compiled kernels exist.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
from typing import Any
|
|
|
|
import tile_kernels.torch as _ref
|
|
|
|
|
|
def __getattr__(name: str) -> Any:
|
|
value = getattr(_ref, name)
|
|
globals()[name] = value
|
|
return value
|
|
|
|
|
|
def __dir__() -> list[str]:
|
|
return sorted(set(globals()) | {n for n in dir(_ref) if not n.startswith("_")})
|