From f05a3ae3e9205f0cfbc5547169bcc02d9c3e73bd Mon Sep 17 00:00:00 2001
From: CalamitousFelicitousness <iowasovereign@gmail.com>
Date: Sat, 11 Jul 2026 09:18:16 +0100
Subject: [PATCH] fix(sdnq): skip compiled dequant for fp8 weights without
 hardware support

Triton has no e4m3 conversions before sm_89, so any compiled graph
touching fp8 storage weights fails with an InductorError on Ampere.
Select the eager dequant and re-quantize paths for e4m3 weights when
the hardware cannot compile them; other dtypes keep compiled dequant.
SDNQ_ALLOW_FP8_COMPILE overrides the detection.
---
 modules/sdnq/common.py      |  8 ++++++++
 modules/sdnq/dequantizer.py | 11 ++++++++---
 2 files changed, 16 insertions(+), 3 deletions(-)

diff --git a/modules/sdnq/common.py b/modules/sdnq/common.py
index 88e5b0482..490ebe89e 100644
--- a/modules/sdnq/common.py
+++ b/modules/sdnq/common.py
@@ -358,6 +358,14 @@ if os.environ.get("SDNQ_ALLOW_FP8_MM", None) is None:
 else:
     is_fp8_mm_supported = os.environ.get("SDNQ_ALLOW_FP8_MM", "0").lower() not in {"0", "false", "no"}
 
+if os.environ.get("SDNQ_ALLOW_FP8_COMPILE", None) is None:
+    if devices.backend == "cuda":
+        is_fp8_compile_supported = bool(torch.cuda.get_device_capability(devices.device) >= (8,9)) # triton has no e4m3 conversions before sm_89
+    else:
+        is_fp8_compile_supported = True
+else:
+    is_fp8_compile_supported = os.environ.get("SDNQ_ALLOW_FP8_COMPILE", "0").lower() not in {"0", "false", "no"}
+
 if os.environ.get("SDNQ_USE_OPENVINO_MM", None) is None:
     use_openvino_mm = bool(devices.backend in {"cpu", "openvino"})
 else:
diff --git a/modules/sdnq/dequantizer.py b/modules/sdnq/dequantizer.py
index b2d2bb0c6..637574168 100644
--- a/modules/sdnq/dequantizer.py
+++ b/modules/sdnq/dequantizer.py
@@ -5,13 +5,17 @@ from dataclasses import dataclass
 import torch
 
 from modules import devices
-from .common import dtype_dict, compile_func, use_contiguous_int8_mm, use_contiguous_fp16_mm, use_tensorwise_fp8_matmul
+from .common import dtype_dict, compile_func, use_contiguous_int8_mm, use_contiguous_fp16_mm, use_tensorwise_fp8_matmul, is_fp8_compile_supported
 from .quant_utils import quantize_int_mm, quantize_uint_mm, quantize_fp_mm, rotate_hadamard, get_hadamard
 from .packed_int import unpack_int
 from .packed_float import unpack_float
 from .layers import SDNQLayer
 
 
+def skip_fp8_compile(weights_dtype: str) -> bool: # triton has no e4m3 conversions before sm_89, compiled dequant would crash
+    return not is_fp8_compile_supported and dtype_dict[weights_dtype]["storage_dtype"] == torch.float8_e4m3fn
+
+
 @devices.inference_context()
 def dequantize_asymmetric(
     weight: torch.Tensor,
@@ -290,7 +294,8 @@ class SDNQDequantizer:
     ) -> tuple[torch.Tensor, torch.FloatTensor]: # pylint: disable=unused-argument
         if hadamard is None and self.use_hadamard and not non_hadamard:
             hadamard = get_hadamard(self.hadamard_group_size, dtype=self.result_dtype, device=weight.device)
-        return re_quantize_matmul_compiled(
+        re_quantize_matmul_func = re_quantize_matmul if skip_fp8_compile(self.weights_dtype) else re_quantize_matmul_compiled
+        return re_quantize_matmul_func(
             self.weights_dtype,
             weight,
             scale,
@@ -322,7 +327,7 @@ class SDNQDequantizer:
         if hadamard is None and self.use_hadamard and not non_hadamard:
             hadamard = get_hadamard(self.hadamard_group_size, dtype=dtype, device=weight.device)
         re_quantize_for_matmul = self.re_quantize_for_matmul or self.is_packed
-        dequantize_weight_func = dequantize_weight if skip_compile else dequantize_weight_compiled
+        dequantize_weight_func = dequantize_weight if skip_compile or skip_fp8_compile(self.weights_dtype) else dequantize_weight_compiled
         return dequantize_weight_func(
             self.weights_dtype,
             weight,