ci: isolate AMX input layout failure and validate contiguous projection

Author: abetlen

.github/scripts/mamba-contiguous.patch

@@ -0,0 +1,13 @@
+diff --git a/src/models/mamba-base.cpp b/src/models/mamba-base.cpp
+index 03ee3805b..9deff6629 100644
+--- a/src/models/mamba-base.cpp
++++ b/src/models/mamba-base.cpp
+@@ -103,7 +103,7 @@ ggml_tensor * llm_build_mamba_base::build_mamba_layer(llm_graph_input_rs * inp,
+         }
+ 
+         // {dt_rank, d_inner} @ {dt_rank, n_seq_tokens, n_seqs} => {d_inner, n_seq_tokens, n_seqs}
+-        dt = build_lora_mm(layer.ssm_dt, dt);
++        dt = build_lora_mm(layer.ssm_dt, ggml_cont(ctx0, dt));
+         dt = ggml_add(ctx0, dt, layer.ssm_dt_b);
+ 
+         cur = x;

.github/scripts/recurrent-diagnostic.patch

@@ -0,0 +1,20 @@
+diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
+index 6862128e6..6e130c76d 100644
+--- a/ggml/src/ggml-backend.cpp
++++ b/ggml/src/ggml-backend.cpp
+@@ -1291,7 +1291,14 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
+         for (int b = 0; b < sched->n_backends && *cur_backend_id == -1; b++) {
+             ggml_backend_sched_set_if_supported(sched, node, b, cur_backend_id);
+         }
+-        GGML_ASSERT(*cur_backend_id != -1);
++        if (*cur_backend_id == -1) {
++            GGML_ABORT("no backend for %s (%s); src0=%s buffer=%s; src1=%s contiguous=%d",
++                    node->name, ggml_op_name(node->op),
++                    node->src[0] ? node->src[0]->name : "none",
++                    node->src[0] && node->src[0]->buffer ? ggml_backend_buffer_name(node->src[0]->buffer) : "none",
++                    node->src[1] ? node->src[1]->name : "none",
++                    node->src[1] ? ggml_is_contiguous(node->src[1]) : 0);
++        }
+     }
+ 
+     // pass 5: split graph, find tensors that need to be copied

.github/workflows/debug-recurrent.yaml

@@ -28,9 +28,18 @@ jobs:
           key: ${{ runner.os }}-model-qwen35-q8-mamba130m-q2-falconh1tiny-q2-bge-small-q4
       - name: CPU details
         run: lscpu
+      - name: Add native diagnostics
+        run: git -C vendor/llama.cpp apply "$GITHUB_WORKSPACE/.github/scripts/recurrent-diagnostic.patch"
       - name: Build
         run: |
           python -m pip install uv
-          python -m uv pip install --system -e '.[all]' --verbose
+          python -m uv pip install --system -e '.[all]' --verbose --config-settings=build-dir=build-diagnostic
       - name: Test with abort diagnostics
+        continue-on-error: true
+        run: python .github/scripts/diagnose-recurrent.py
+      - name: Test contiguous projection input fix
+        run: |
+          git -C vendor/llama.cpp apply "$GITHUB_WORKSPACE/.github/scripts/mamba-contiguous.patch"
+          python -m uv pip install --system --reinstall-package llama-cpp-python -e '.[all]' --verbose --config-settings=build-dir=build-diagnostic
+      - name: Verify fixed model
         run: python .github/scripts/diagnose-recurrent.py