DeepSeek-V4 B2: block-FP8 / packed-FP4 weight-loading + expert-emission primitives (DRAFT)

src/mobius/_configs/_quantization.py

@@ -74,11 +74,43 @@ def from_transformers(cls, hf_config) -> QuantizationConfig | None:
                 "onnxruntime_USE_FP4_QMOE=ON). Export the unquantized (bf16) "
                 "checkpoint instead, or quantize the bf16 export via Olive."
             )
+        # Block-scaled fp8 (E4M3 weight + 2D UE8M0 block scale) and packed-fp4
+        # routed experts (I8-packed E2M1 nibbles + UE8M0 micro-scale) are a
+        # mixed-precision layout this INT4/per-tensor path cannot load — the
+        # packed [out, in/2] fp4 expert vs its logical [out, in] initializer
+        # produces a confusing "Weight shape mismatch". Detect it by property
+        # (not model name, not the ``quant_method`` string) and fail closed with
+        # a typed, actionable blocker naming the real layout + the runtime ABI
+        # gap. Checked before the ``quant_method == "none"`` early-return because
+        # a checkpoint can advertise fp4 experts via top-level ``expert_dtype``
+        # while leaving ``quant_method`` unset.
+        from mobius.integrations._block_quant import BlockQuantScheme
+
+        scheme = BlockQuantScheme.from_quantization_config(
+            qc, expert_dtype=getattr(hf_config, "expert_dtype", None)
+        )
+        if scheme is not None:
+            from mobius.integrations._block_quant import BlockQuantExportError
+
+            raise BlockQuantExportError(
+                "Block-scaled FP8 / packed-FP4 checkpoint is not loadable by the "
+                "INT4/per-tensor quantization path. Detected "
+                f"quant_method={scheme.quant_method!r}, "
+                f"weight_block_size={list(scheme.weight_block_size) or None}, "
+                f"expert_dtype={scheme.expert_dtype!r}: block-FP8 projections "
+                "(E4M3 weight + 2D UE8M0 block scale) and/or FP4-packed routed "
+                "experts (I8-packed E2M1 nibbles + UE8M0 micro-scale). Parse and "
+                "validate these by property with mobius.integrations._block_quant "
+                "(BlockQuantScheme / classify_tensor / QuantizedTensorDescriptor); "
+                "the routed-expert emission gate (plan_routed_expert_bank) reports "
+                "the exact onnx-genai nxrt ABI gap. Native export is blocked until "
+                "the runtime gains a block-FP8 / planar-FP4 BlockFormat."
+            )
         if method == "none":
             return None
-        # FP8 per-tensor quantization (float8_e4m3fn + scalar scale)
-        # is handled by dtype casting in _assign_weight(), not by
-        # QuantizedLinear block quantization.
+        # Per-tensor fp8 (float8_e4m3fn + a scalar scale) is handled by dtype
+        # casting in _assign_weight(), so it returns None here. (Block-scaled
+        # fp8 was already routed to the typed blocker above.)
         if method == "fp8":
             return None
         return cls(