liangsu9988 commited on
Commit
f556a68
·
1 Parent(s): d3aaf4c

Document NVFP4 and experimental INT4RHT tiers

Browse files
Files changed (1) hide show
  1. README.md +31 -3
README.md CHANGED
@@ -6,10 +6,38 @@ Apache-2.0 SageAttention3 implementation in `thu-ml/SageAttention`.
6
  Use this package as an explicitly gated lower-precision tier for long video or
7
  audio sequences. Its preferred `sage3_prefill_fp4_bf16` entry accepts raw BF16
8
  NHD Q/K/V and owns Sage3 centering, padding, FP4 quantization, and correction
9
- preparation behind a caller-owned CUDA Graph-safe workspace. Head dimensions
10
- 64 and 128 are supported. It does not support GQA/MQA in v1 and does not replace
 
 
11
  the higher-fidelity
12
  `flashrt/sageattention2-blackwell` default.
13
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
14
  See `README.md` for preprocessing, API usage, accuracy expectations, and the
15
- complete capability contract.
 
6
  Use this package as an explicitly gated lower-precision tier for long video or
7
  audio sequences. Its preferred `sage3_prefill_fp4_bf16` entry accepts raw BF16
8
  NHD Q/K/V and owns Sage3 centering, padding, FP4 quantization, and correction
9
+ preparation behind a caller-owned CUDA Graph-safe workspace. Head dimension 64
10
+ is available in CUDA 12.8+ artifacts; head dimension 128 requires a CUDA 13.0+
11
+ artifact. Query `capabilities()` rather than assuming a head-dimension set. It
12
+ does not support GQA/MQA in v1 and does not replace
13
  the higher-fidelity
14
  `flashrt/sageattention2-blackwell` default.
15
 
16
+ ## Published and experimental variants
17
+
18
+ The v1 artifacts currently publish one executable quantization tier:
19
+
20
+ - **Sage3 NVFP4/E2M1**: FP4 Q/K/V preparation plus block-scaled FP4
21
+ attention, exposed through `sage3_prefill_fp4_bf16` and the lower-level
22
+ `prepare_qkv_fp4_nhd` / `blockscaled_fp4_attention_static` APIs.
23
+
24
+ The **INT4RHT** tier is not present in the v1 binaries and is not callable yet.
25
+ It is a design-stage fallback for deployments that need an accuracy point
26
+ between SageAttention2 and Sage3 NVFP4. Its proposed contract is symmetric
27
+ INT4 Q/K with per-16-element floating-point scales, a shared Hadamard-128
28
+ rotation after RoPE and centering, and the SageAttention2 FP8 P/V path:
29
+
30
+ ```python
31
+ out = sage3.sage3_prefill_int4rht_bf16(
32
+ q, k, v, out=out_buffer, workspace=workspace,
33
+ )
34
+ ```
35
+
36
+ That symbol will only be published after it passes all three gates: cosine at
37
+ least 0.9999 against the INT4RHT reference on random and real activations,
38
+ model-capture cosine at least 0.997, and all-in latency below SageAttention2
39
+ FP8V. Until then, callers must not probe or advertise it as a supported
40
+ variant.
41
+
42
  See `README.md` for preprocessing, API usage, accuracy expectations, and the
43
+ complete capability contract.