fix: clip 模型使用 non-mask softmax

Signed-off-by: YdrMaster <[email protected]>

Author: YdrMaster

Cargo.toml

@@ -38,7 +38,7 @@ itertools = "0.13"
 env_logger = "0.11"
 build-script-cfg = "0.0"
 
-operators = { git = "https://github.com/YdrMaster/operators-rs", rev = "f872f7e", default-features = false }
+operators = { git = "https://github.com/YdrMaster/operators-rs", rev = "61789f7", default-features = false }
 
 search-cl-tools = { git = "https://github.com/InfiniTensor/clrt", rev = "f69b160" }
 search-infini-tools = { git = "https://github.com/InfiniTensor/infini-rt", rev = "e8362c3" }

models/clip/common/src/compute.rs

@@ -5,6 +5,7 @@ use operators::{
     add_rows::{self, AddRows},
     attention::{self, Attention},
     conv::{self, Conv},
+    fuesd_softmax::AttnMask,
     gelu::{self, Gelu},
     layer_norm::{self, LayerNorm},
     mat_mul::{self, MatMul},
@@ -422,6 +423,7 @@ where
                 v_base: v.base(),
                 o_layout: o.layout(),
                 o_base: o.base_mut(),
+                mask: AttnMask::None,
             },
             workspace,
             queue_alloc,

models/gpt2/common/src/compute.rs

@@ -5,6 +5,7 @@ use operators::{
     add_rows::{self, AddRows},
     all_reduce::{self, AllReduce, ReduceOp},
     attention_kv_cached::{self, AttnKVCached},
+    fuesd_softmax::AttnMask,
     gelu::{self, Gelu},
     layer_norm::{self, LayerNorm},
     mat_mul::{self, MatMul},
@@ -399,6 +400,7 @@ where
                 k_cache_base: kc.base_mut(),
                 v_cache_layout: vc.layout(),
                 v_cache_base: vc.base_mut(),
+                mask: AttnMask::Causal,
                 pos: pos.into(),
             },
             workspace,

models/llama/common/src/compute.rs

@@ -7,6 +7,7 @@ use itertools::izip;
 use operators::{
     all_reduce::{self, AllReduce, ReduceOp},
     attention_kv_cached::{self, AttnKVCached},
+    fuesd_softmax::AttnMask,
     mat_mul::{self, MatMul},
     rearrange::{self, Rearrange},
     rms_norm::{self, RmsNorm},
@@ -520,6 +521,7 @@ where
                 k_cache_base: kc.base_mut(),
                 v_cache_layout: vc.layout(),
                 v_cache_base: vc.base_mut(),
+                mask: AttnMask::Causal,
                 pos: pos.into(),
             },
             workspace,