Fix ultra-long context prefilling in Qwen3 MoE GGUF models

Author: guoqingbao

Cargo.toml

@@ -27,10 +27,10 @@ license = "MIT"
 rust-version = "1.82"
 
 [workspace.dependencies]
-candle-core = { git = "https://github.com/EricLBuehler/candle.git", version = "0.9.1", rev = "5e6c385" }
-candle-nn = { git = "https://github.com/EricLBuehler/candle.git", version = "0.9.1", rev = "5e6c385" }
-candle-flash-attn-v3 = { git = "https://github.com/EricLBuehler/candle.git", version = "0.9.1", rev = "5e6c385" }
-candle-flash-attn = { git = "https://github.com/EricLBuehler/candle.git", version = "0.9.1", rev = "5e6c385" }
+candle-core = { path = "../mistralrs/candle/candle-core", version = "0.9.1" }
+candle-nn = { path = "../mistralrs/candle/candle-nn", version = "0.9.1" }
+candle-flash-attn-v3 = { path = "../mistralrs/candle/candle-flash-attn-v3", version = "0.9.1"}
+candle-flash-attn = { path = "../mistralrs/candle/candle-flash-attn", version = "0.9.1" }
 # candle-core = { path = "../candle/candle-core" }
 # candle-nn = { path = "../candle/candle-nn" }
 # candle-flash-attn-v3 = { path = "../candle/candle-flash-attn-v3" }