mirror of
https://github.com/deepseek-ai/DeepInfer.git
synced 2026-11-01 17:26:25 +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.
98 lines
3.3 KiB
Python
98 lines
3.3 KiB
Python
"""`tile_kernels.transform` -- rotary position embedding.
|
|
|
|
This is the widest gap between the two signatures, and the one worth reading carefully:
|
|
|
|
* the compiled kernel takes `cos_sin_cache`, a single tensor of ``cat([cos, sin], -1)``, and
|
|
**rotates in place**, returning ``None``;
|
|
* the reference takes `cos` and `sin` as separate tensors and **returns a new tensor**.
|
|
|
|
So the adapter has to split the cache, call the reference, and write the result back into the
|
|
caller's tensor -- because the call site relies on the in-place contract
|
|
(`rotary_embedding.py` documents it and discards the return value).
|
|
|
|
Measured against the compiled kernel: identical except one bf16 last-bit tie per 65536 elements,
|
|
which is rounding, not a logic difference.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import torch
|
|
|
|
from deepinfer.tk_compat import use_references
|
|
|
|
if not use_references():
|
|
from tile_kernels import transform as _compiled
|
|
|
|
from tile_kernels import torch as _ref
|
|
|
|
|
|
def apply_rotary(
|
|
query: torch.Tensor,
|
|
cos_sin_cache: torch.Tensor,
|
|
key: torch.Tensor | None = None,
|
|
positions: torch.Tensor | None = None,
|
|
interleaved: bool = False,
|
|
conjugate: bool = False,
|
|
seqlen_offset: int = 0,
|
|
) -> None:
|
|
"""Rotate `query` (and `key`) in place. Returns None, like the compiled kernel."""
|
|
if not use_references():
|
|
return _compiled.apply_rotary(
|
|
query,
|
|
cos_sin_cache,
|
|
key,
|
|
positions,
|
|
interleaved,
|
|
conjugate,
|
|
seqlen_offset,
|
|
)
|
|
|
|
# The cache stores cos and sin halves side by side; rot_dim is their total.
|
|
rot_dim = cos_sin_cache.shape[-1]
|
|
if rot_dim % 2:
|
|
raise ValueError(f"cos_sin_cache last dim must be even, got {rot_dim}")
|
|
half = rot_dim // 2
|
|
cos = cos_sin_cache[..., :half]
|
|
sin = cos_sin_cache[..., half:]
|
|
|
|
def rotate(x: torch.Tensor) -> None:
|
|
# The reference wants (seqlen, heads, rot_dim) or 4-D with a batch axis; the caller passes
|
|
# (num_tokens, num_heads, rot_dim). Cast for the reshape and back after.
|
|
original_shape = x.shape
|
|
if x.dim() == 3:
|
|
seqlen, heads, dim = original_shape
|
|
view = x.view(seqlen, heads, dim)
|
|
elif x.dim() == 4:
|
|
batch, seqlen, heads, dim = original_shape
|
|
view = x.view(batch * seqlen, heads, dim)
|
|
else:
|
|
raise ValueError(f"apply_rotary expects 3-D or 4-D input, got {x.dim()}-D")
|
|
|
|
# `positions` is per-token; a flattened batch axis would need them tiled, which the
|
|
# compiled kernel also does not accept.
|
|
if x.dim() == 4 and positions is not None:
|
|
raise NotImplementedError(
|
|
"4-D query with explicit positions has no reference equivalent: the reference "
|
|
"indexes positions along its batch axis. Pass None to use the default ordering."
|
|
)
|
|
|
|
rotated = _ref.rotary_embedding_ref(
|
|
view,
|
|
cos,
|
|
sin,
|
|
positions,
|
|
seqlen_offsets=seqlen_offset,
|
|
interleaved=interleaved,
|
|
conjugate=conjugate,
|
|
)
|
|
# In-place contract: the caller keeps using its own tensor.
|
|
x.copy_(rotated.view(original_shape))
|
|
|
|
rotate(query)
|
|
if key is not None:
|
|
rotate(key)
|
|
|
|
|
|
def backend() -> str:
|
|
return "ref" if use_references() else "cuda"
|