Restrict rocMLIR fusion to non-RDNA GPUs on released rocMLIR

migraphx 10.1 force-routes all fused conv/dot ops (and convolution_
backwards, and kv-cache attention) through rocMLIR on gfx11/gfx12
("navi"), but released rocMLIR (<= 7.2.4) has no RDNA WMMA rock
kernels: RDNA support exists only on rocMLIR main (gfx1250 wmma work,
late 2025) and was never tagged. Against a released rocMLIR every such
fusion dies at compile time with:

  compile_ops.cpp:433: add_compiles: No solutions provided for gpu::mlir_op

AMD's ROCm 10.1 distribution pairs migraphx 10.1 with an unreleased
rocMLIR snapshot that carries the RDNA support; that combination is not
reproducible from tags.

Default navi to mlir_mode::none so conv/dot fall back to MIOpen and
rocBLAS; MIGRAPHX_MLIR_USE_SPECIFIC_OPS still forces ops on for
testing. The gfx12 convolution_backwards forcing is commented out
(accepting the less accurate MIOpen half path) and kv-cache attention
is gated off on navi.

DROP THIS PATCH when rocMLIR tags a release with RDNA WMMA rock
support (watch for a tag after 7.2.4 containing the gfx1250 wmma
commits).
--- a/src/targets/gpu/fuse_mlir.cpp
+++ b/src/targets/gpu/fuse_mlir.cpp
@@ -1645,8 +1645,16 @@
             return mlir_mode::none;
         if(env_op_usage().is_requested(option))
             return mlir_mode::all;
+        // Released rocMLIR (<= 7.2.4) has no RDNA WMMA rock kernels; RDNA
+        // support exists only on rocMLIR main (gfx1250 wmma work, late
+        // 2025, unreleased). Forcing mlir_mode::all on navi against a
+        // released rocMLIR makes every fused conv/dot fail with
+        // "No solutions provided for gpu::mlir_op". Default to none on
+        // navi instead; MIGRAPHX_MLIR_USE_SPECIFIC_OPS (checked above)
+        // can still force ops on for testing. Drop this hunk when rocMLIR
+        // tags RDNA rock support.
         if(is_navi)
-            return mlir_mode::all;
+            return mlir_mode::none;
 #if !MIGRAPHX_USE_MIOPEN
         if(contains(option, "conv"))
             return mlir_mode::all;
@@ -1667,7 +1675,9 @@
     match::find_matches(mpm, find_channel_slice_convolution{});
     mpm.run_pass(dead_code_elimination{});
 
-    match::find_matches(mpm, find_mlir_kv_cache_attention_op{mlir_mode::all});
+    match::find_matches(mpm,
+                        find_mlir_kv_cache_attention_op{.dot_mode = is_navi ? mlir_mode::none
+                                                                                : mlir_mode::all});
     mpm.run_pass(dead_code_elimination{});
 
     match::find_matches(mpm, find_mlir_attention_op{});
@@ -1683,10 +1693,13 @@
     // env-var overrides.
     mlir_mode conv_backwards_mode =
         get_mode("convolution_backwards", MIGRAPHX_USE_MIOPEN ? mlir_mode::none : mlir_mode::all);
-    if(starts_with(device_name, "gfx12"))
-    {
-        conv_backwards_mode = mlir_mode::all;
-    }
+    // Same rocMLIR RDNA gap as above: gfx12 would be forced through
+    // rocMLIR for convolution_backwards, which crashes with no solutions
+    // against released rocMLIR. Accept the less accurate MIOpen half
+    // path instead. Restore the gfx12 forcing below when rocMLIR tags
+    // RDNA rock support:
+    //   if(starts_with(device_name, "gfx12"))
+    //       conv_backwards_mode = mlir_mode::all;
 
     match::find_matches(
         mpm,