[Draft · diff only] Emulator, stage tests and measurements for the static Mamba/KDA path

.github/workflows/transactional_emulator.yml

@@ -33,7 +33,13 @@ jobs:
   rust-lint:
     runs-on: ubuntu-latest
     steps:
+      # Submodules: clippy compiles `--all-targets`, and one test guard reads the
+      # compiler's `doc/operation.svh` through `include_str!` to check that every
+      # opcode the compiler emits decodes here. Without the submodule that is a
+      # compile error, not a skipped test.
       - uses: actions/checkout@v4
+        with:
+          submodules: recursive
 
       # Pinned: with `-D warnings`, a rolling channel reds unrelated PRs.
       - name: Install Rust toolchain
@@ -147,3 +153,24 @@ jobs:
       - name: Run emulator timing gates
         run: |
           nix develop --command bash -c "just test-timing-gates"
+
+      # Mamba-2 / selective SSM. Synthetic like the tests above -- no checkpoint,
+      # no HF libs. Inputs are drawn exactly representable in MX-FP8 so the
+      # reported error is the kernel's rather than the input quantisation's.
+      #
+      # All four cases are bit-exact (max absolute error 0.000000), so any
+      # regression shows up as a hard failure rather than as tolerance drift.
+      - name: Run Mamba-2 stage tests
+        run: |
+          nix develop --command bash -c "just test-mamba2-all"
+
+      # The chunked-SSD reference vs the plain recurrence, plus the cross-repo
+      # stage-marker contract. Pure torch.
+      - name: Run Mamba-2 reference cross-check
+        run: |
+          nix develop --command bash -c "just test-mamba2-reference"
+
+      # Bandwidth term + the two KV-cost bugs it exposed in the analytic model.
+      - name: Run analytic performance-model regression tests
+        run: |
+          nix develop --command bash -c "just test-perf-model"

analytic_models/__init__.py

@@ -1,3 +1,10 @@
-from . import memory, performance, utilisation
+"""PLENA analytic models.
 
-__all__ = ["memory", "performance", "utilisation"]
+Only the `performance` package exists in this checkout; the previous
+`from . import memory, performance, utilisation` raised ImportError on any
+`import analytic_models` because `memory` and `utilisation` were never added.
+"""
+
+from . import performance
+
+__all__ = ["performance"]

analytic_models/performance/customISA_lib.json

@@ -93,6 +93,10 @@
         "alone" : "6 + 3 + VECTOR_MUL_CYCLES + 1",
         "pipelined" : "VECTOR_MUL_CYCLES"
     },
+    "V_FMA_VF" : {
+        "alone" : "6 + 3 + VECTOR_MUL_CYCLES + 1",
+        "pipelined" : "VECTOR_MUL_CYCLES"
+    },
 
     "V_EXP_V" : {
         "alone" : "6 + 3 + VECTOR_EXP_CYCLES + 1",
@@ -119,6 +123,31 @@
         "pipelined" : " 1 + VECTOR_MAX_CYCLES"
     },
 
+    "V_MAX_VF" : {
+        "alone" : "6 + 3 + VECTOR_MAX_CYCLES + 1",
+        "pipelined" : "1 + VECTOR_MAX_CYCLES"
+    },
+
+    "V_MIN_VF" : {
+        "alone" : "6 + 3 + VECTOR_MAX_CYCLES + 1",
+        "pipelined" : "1 + VECTOR_MAX_CYCLES"
+    },
+
+    "V_PS_V" : {
+        "alone" : "6 + 3 + VECTOR_PREFIX_SCAN_CYCLES + 1",
+        "pipelined" : "VECTOR_PREFIX_SCAN_CYCLES"
+    },
+
+    "V_SHFT_V" : {
+        "alone" : "6 + 3 + VECTOR_SHIFT_CYCLES + 1",
+        "pipelined" : "VECTOR_SHIFT_CYCLES"
+    },
+
+    "V_SOFTPLUS_V" : {
+        "alone" : "6 + 3 + VECTOR_SOFTPLUS_CYCLES + 1",
+        "pipelined" : "1 + VECTOR_SOFTPLUS_CYCLES"
+    },
+
     "S_BASIC" : {
         "alone" : "   SCALAR_FP_BASIC_CYCLES + 1",
         "pipelined" : "SCALAR_FP_BASIC_CYCLES"
@@ -174,6 +203,11 @@
         "pipelined" : "1"
     },
 
+    "S_MAP_FP_V" : {
+        "alone" : "6 + 1",
+        "pipelined" : "1"
+    },
+
     "S_ADD_INT" : {
         "alone" : "2 + SCALAR_INT_BASIC_CYCLES + 1",
         "pipelined" : "SCALAR_INT_BASIC_CYCLES"