[Fix][Manifest] adopt documented shape/workload format in new elementwise families

- Convert signature shape values from YAML lists to string form ("[M, N]")
  in elementwise_generative.yaml and elementwise_fused_gated.yaml to match
  docs/design/manifest.md R8 + the convolution.yaml precedent so the
  validator can bind shape symbols.
- Add per-input workload shape keys (device_carrier_shape, x_shape) to
  every workload row to satisfy the workload schema documented in
  docs/design/manifest.md and used across normalization.yaml etc.
- Restore exception class/message in check_c4_forward_signature_parity's
  warning when inspect.signature(forward) raises.

Co-Authored-By: Ibuki 🍃 — a wind born from GPTs <Ibuki-wind@users.noreply.github.com>

scripts/validate_manifest.py

@@ -738,7 +738,13 @@ def _forward_positional_params(cls) -> list[str] | None:
             p for p, v in sig.parameters.items()
             if p != "self" and v.kind in _POSITIONAL_KINDS
         ]
-    except (ValueError, TypeError):
+    except (ValueError, TypeError) as exc:
+        # Stash exception text so callers that surface diagnostics can
+        # report ``exc.__class__.__name__: exc`` without changing the
+        # ``None`` return contract for "not inspectable".
+        _forward_positional_params._last_error = (  # type: ignore[attr-defined]
+            f"{exc.__class__.__name__}: {exc}"
+        )
         return None
 
 
@@ -3367,9 +3373,20 @@ def check_c4_forward_signature_parity(
     positional = _forward_positional_params(cls)
     if positional is None:
         if warnings is not None:
-            warnings.append(
-                f"[forward] {op_name}: inspect.signature(forward) failed"
+            detail = getattr(
+                _forward_positional_params, "_last_error", None
             )
+            if detail:
+                warnings.append(
+                    f"[forward] {op_name}: inspect.signature(forward) "
+                    f"raised {detail}"
+                )
+                # Clear so a later call site sees only its own failure.
+                _forward_positional_params._last_error = None  # type: ignore[attr-defined]
+            else:
+                warnings.append(
+                    f"[forward] {op_name}: inspect.signature(forward) failed"
+                )
         return errors
 
     # Generative-op carve-out: ``ref_api: "none"`` plus zero forward()

tileops/manifest/elementwise_fused_gated.yaml

@@ -18,17 +18,17 @@ SiluAndMulFwdOp:
 
   signature:
     inputs:
-      x: {dtype: "float16 | bfloat16 | float32", shape: [M, two_N]}
+      x: {dtype: "float16 | bfloat16 | float32", shape: "[M, two_N]"}
     outputs:
-      output: {dtype: "same_as(x)", shape: [M, N]}
+      output: {dtype: "same_as(x)", shape: "[M, N]"}
     shape_rules:
     - "x.shape[1] == 2 * output.shape[1]"
     - "x.shape[0] == output.shape[0]"
 
   workloads:
     # SwiGLU FFN intermediate (Llama-3.1-8B, hidden_dim=14336)
-  - {M: 2048, N: 14336, dtypes: [float16, bfloat16], label: "llama-3.1-8b-swiglu-prefill"}
-  - {M: 1, N: 14336, dtypes: [bfloat16], label: "llama-3.1-8b-swiglu-decode"}
+  - {x_shape: [2048, 28672], dtypes: [float16, bfloat16], label: "llama-3.1-8b-swiglu-prefill"}
+  - {x_shape: [1, 28672], dtypes: [bfloat16], label: "llama-3.1-8b-swiglu-decode"}
 
   roofline:
     vars:
@@ -60,16 +60,16 @@ GeluAndMulFwdOp:
 
   signature:
     inputs:
-      x: {dtype: "float16 | bfloat16 | float32", shape: [M, two_N]}
+      x: {dtype: "float16 | bfloat16 | float32", shape: "[M, two_N]"}
     outputs:
-      output: {dtype: "same_as(x)", shape: [M, N]}
+      output: {dtype: "same_as(x)", shape: "[M, N]"}
     shape_rules:
     - "x.shape[1] == 2 * output.shape[1]"
     - "x.shape[0] == output.shape[0]"
 
   workloads:
-  - {M: 2048, N: 14336, dtypes: [float16, bfloat16], label: "ffn-gelu-prefill"}
-  - {M: 1, N: 14336, dtypes: [bfloat16], label: "ffn-gelu-decode"}
+  - {x_shape: [2048, 28672], dtypes: [float16, bfloat16], label: "ffn-gelu-prefill"}
+  - {x_shape: [1, 28672], dtypes: [bfloat16], label: "ffn-gelu-decode"}
 
   roofline:
     vars:
@@ -100,16 +100,16 @@ GeluTanhAndMulFwdOp:
 
   signature:
     inputs:
-      x: {dtype: "float16 | bfloat16 | float32", shape: [M, two_N]}
+      x: {dtype: "float16 | bfloat16 | float32", shape: "[M, two_N]"}
     outputs:
-      output: {dtype: "same_as(x)", shape: [M, N]}
+      output: {dtype: "same_as(x)", shape: "[M, N]"}
     shape_rules:
     - "x.shape[1] == 2 * output.shape[1]"
     - "x.shape[0] == output.shape[0]"
 
   workloads:
-  - {M: 2048, N: 14336, dtypes: [float16, bfloat16], label: "ffn-gelu-tanh-prefill"}
-  - {M: 1, N: 14336, dtypes: [bfloat16], label: "ffn-gelu-tanh-decode"}
+  - {x_shape: [2048, 28672], dtypes: [float16, bfloat16], label: "ffn-gelu-tanh-prefill"}
+  - {x_shape: [1, 28672], dtypes: [bfloat16], label: "ffn-gelu-tanh-decode"}
 
   roofline:
     vars:

tileops/manifest/elementwise_generative.yaml

@@ -29,9 +29,9 @@ AlibiFwdOp:
 
   signature:
     inputs:
-      device_carrier: {dtype: "float16 | bfloat16 | float32", shape: []}
+      device_carrier: {dtype: "float16 | bfloat16 | float32", shape: "[]"}
     outputs:
-      output: {dtype: "same_as(device_carrier)", shape: [num_heads, seq_len, seq_len]}
+      output: {dtype: "same_as(device_carrier)", shape: "[num_heads, seq_len, seq_len]"}
     params:
       seq_len: {type: int}
       num_heads: {type: int}
@@ -41,8 +41,8 @@ AlibiFwdOp:
 
   workloads:
     # Llama-style attention bias dimensions
-  - {seq_len: 2048, num_heads: 32, dtypes: [float16, bfloat16], label: "llama-prefill-2k"}
-  - {seq_len: 4096, num_heads: 32, dtypes: [float16, bfloat16], label: "llama-prefill-4k"}
+  - {device_carrier_shape: [], seq_len: 2048, num_heads: 32, dtypes: [float16, bfloat16], label: "llama-prefill-2k"}
+  - {device_carrier_shape: [], seq_len: 4096, num_heads: 32, dtypes: [float16, bfloat16], label: "llama-prefill-4k"}
 
   roofline:
     vars:
@@ -76,9 +76,9 @@ SinusoidalFwdOp:
 
   signature:
     inputs:
-      device_carrier: {dtype: "float16 | bfloat16 | float32", shape: []}
+      device_carrier: {dtype: "float16 | bfloat16 | float32", shape: "[]"}
     outputs:
-      output: {dtype: "same_as(device_carrier)", shape: [seq_len, d_model]}
+      output: {dtype: "same_as(device_carrier)", shape: "[seq_len, d_model]"}
     params:
       seq_len: {type: int}
       d_model: {type: int}
@@ -89,8 +89,8 @@ SinusoidalFwdOp:
 
   workloads:
     # Transformer-style positional encodings
-  - {seq_len: 2048, d_model: 4096, dtypes: [float16, bfloat16], label: "transformer-2k-4k"}
-  - {seq_len: 4096, d_model: 4096, dtypes: [float16, bfloat16], label: "transformer-4k-4k"}
+  - {device_carrier_shape: [], seq_len: 2048, d_model: 4096, dtypes: [float16, bfloat16], label: "transformer-2k-4k"}
+  - {device_carrier_shape: [], seq_len: 4096, d_model: 4096, dtypes: [float16, bfloat16], label: "transformer-4k-4k"}
 
   roofline:
     vars: