Fix nvfp4 scaling issue (#456)

Author: guoqingbao

Cargo.toml

@@ -45,7 +45,7 @@ dirs = "5.0.1"
 minijinja = { version = "2.10.2", features = ["builtins", "json"] }
 minijinja-contrib = { version = "2.10.2", features = ["pycompat"] }
 thiserror = "1.0.58"
-attention-rs = { git = "https://github.com/guoqingbao/attention.rs.git", version="0.6.6", rev = "2e0db42" }
+attention-rs = { git = "https://github.com/guoqingbao/attention.rs.git", version="0.6.7", rev = "70ed0d0" }
 metal = { version = "0.27.0", features = ["mps"], optional = true }
 lazy_static = {version = "1.4.0"}
 interprocess = "2.2.2"

src/backend/graph.rs

@@ -620,6 +620,8 @@ impl<M: CudaGraphModule> GraphCapturer<M> {
                     mamba_slot_mapping: Some(mamba_slot_mapping.narrow(0, 0, bs)?),
                     slot_mapping: slot_mapping.narrow(0, 0, bs)?,
                     block_tables: Some(block_tables.narrow(0, 0, bs)?),
+                    block_tables_host: None,
+                    context_lens_host: None,
                     context_lens: Some(context_lens.narrow(0, 0, bs)?),
                     cu_seqlens_q: None,
                     cu_seqlens_k: None,
@@ -793,6 +795,8 @@ impl<M: CudaGraphModule> GraphCapturer<M> {
             mamba_slot_mapping: Some(mamba_slot_mapping.clone()),
             slot_mapping: slot_mapping.clone(),
             block_tables: Some(block_tables.clone()),
+            block_tables_host: None,
+            context_lens_host: None,
             context_lens: Some(context_lens.clone()),
             cu_seqlens_q: Some(cu_seqlens_q.clone()),
             cu_seqlens_k: Some(cu_seqlens_k.clone()),