[neilmsft]

Adds support for per channel quantized weight export and configurable MatMulNBits accuracy levels through the builder APIs. Supports explicit accuracy levels from 0 through 4 without changing weight precision, while preserving existing defaults when no override is provided.

[github-actions]

Performance Comparison

Comparing 2e5508f1 → 06e00978

Model	Metric	Baseline	Current	Delta
bert (feature-extraction)	model_size_bytes	359 KB	359 KB	+0.0%	⚪
bert (feature-extraction)	num_nodes	68	68	+0.0%	⚪
falcon	model_size_bytes	364 KB	364 KB	+0.0%	⚪
falcon	num_nodes	66	66	+0.0%	⚪
gemma2	model_size_bytes	428 KB	428 KB	+0.0%	⚪
gemma2	num_nodes	105	105	+0.0%	⚪
gpt2	model_size_bytes	324 KB	324 KB	+0.0%	⚪
gpt2	num_nodes	54	54	+0.0%	⚪
llama	model_size_bytes	425 KB	425 KB	+0.0%	⚪
llama	num_nodes	60	60	+0.0%	⚪
llama (static-cache)	model_size_bytes	425 KB	425 KB	+0.0%	⚪
llama (static-cache)	num_nodes	56	56	+0.0%	⚪
mamba (ssm-text-generation)	model_size_bytes	296 KB	296 KB	+0.0%	⚪
mamba (ssm-text-generation)	num_nodes	94	94	+0.0%	⚪
phi3	model_size_bytes	421 KB	421 KB	+0.0%	⚪
phi3	num_nodes	58	58	+0.0%	⚪
phi3 (static-cache)	model_size_bytes	421 KB	421 KB	+0.0%	⚪
phi3 (static-cache)	num_nodes	54	54	+0.0%	⚪
qwen2	model_size_bytes	425 KB	425 KB	+0.0%	⚪
qwen2	num_nodes	60	60	+0.0%	⚪
qwen2 (static-cache)	model_size_bytes	425 KB	425 KB	+0.0%	⚪
qwen2 (static-cache)	num_nodes	56	56	+0.0%	⚪
qwen3_5_moe (hybrid-text-generation)	model_size_bytes	506 KB	506 KB	-0.0%	⚪
qwen3_5_moe (hybrid-text-generation)	num_nodes	269	265	-1.5%	⚪
qwen3_5_text (hybrid-text-generation)	model_size_bytes	458 KB	458 KB	-0.0%	⚪
qwen3_5_text (hybrid-text-generation)	num_nodes	131	127	-3.1%	🟢
qwen3_5_vl (hybrid-qwen-vl)	model_size_bytes	977 KB	977 KB	-0.0%	⚪
qwen3_5_vl (hybrid-qwen-vl)	num_nodes	461	450	-2.4%	🟢
t5 (seq2seq)	model_size_bytes	836 KB	836 KB	+0.0%	⚪
t5 (seq2seq)	num_nodes	176	176	+0.0%	⚪
whisper (speech-to-text)	model_size_bytes	1008 KB	1008 KB	+0.0%	⚪
whisper (speech-to-text)	num_nodes	128	128	+0.0%	⚪

No performance regressions.

[Copilot]

🟡 Changes recommended

Global per-channel loading and explicit full-channel group sizes still have normalization mismatches that can prevent valid checkpoints from loading.

2 open issues

• Fix global per-channel sidecar normalization for leaf projections · New
• Handle explicit full-channel block sizes in weight preprocessing · New

What changed in this PR

Adds per-channel quantized weight export and configurable MatMulNBits accuracy levels to mobius’s graph-building APIs, preserving existing defaults when no override is supplied.

Changes:

• Threads accuracy levels 0–4 through core, Transformers, and Diffusers builders.
• Adds full-channel packed-weight layouts and padding-aware normalization.
• Extends tests for layouts, numerical parity, option forwarding, and context isolation.

File	Description
src/​mobius/​weights/​_codecs.py	Passes logical input width to weight normalization.
src/​mobius/​models/​gemma4_test.py	Tests per-channel decoder layout isolation.
src/​mobius/​integrations/​transformers/​_builder.py	Validates and forwards accuracy overrides.
src/​mobius/​integrations/​transformers/​_builder_test.py	Tests forwarding and unsupported-builder rejection.
src/​mobius/​integrations/​diffusers/​_builder.py	Forwards accuracy overrides to component builds.
src/​mobius/​integrations/​diffusers/​_builder_test.py	Tests accuracy-option forwarding.
src/​mobius/​functions/​matmul_nbits.py	Trims packed padding before dequantization.
src/​mobius/​functions/​matmul_nbits_test.py	Tests padded-weight numerical correctness.
src/​mobius/​components/​_quantized_linear.py	Supports full-channel layouts and accuracy overrides.
src/​mobius/​components/​_quantized_linear_test.py	Tests layouts, accuracy attributes, and export parity.
src/​mobius/​_weight_utils.py	Adds per-channel packed-weight normalization.
src/​mobius/​_weight_utils_test.py	Tests packing preservation and layout validation.
src/​mobius/​_quantization_utils.py	Resolves per-channel block size from input width.
src/​mobius/​_component_quantization.py	Matches resolved layouts and supplies logical widths.
src/​mobius/​_component_quantization_test.py	Tests component-specific per-channel normalization.
src/​mobius/​_builder.py	Exposes the core accuracy override.
src/​mobius/​_builder_test.py	Tests exported accuracy attributes.
src/​mobius/​_build_context.py	Stores and validates scoped accuracy overrides.
src/​mobius/​_build_context_test.py	Tests validation, restoration, and concurrent isolation.

_{Review effort: Balanced}

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[github-actions]

🏗️ Architecture Diff

Comparing 2e5508f1 → 06e00978

Model	Sub-model	Changes	Status
bert (feature-extraction)	model	0	⚪
falcon	model	0	⚪
gemma2	model	0	⚪
gemma4 (gemma4)	decoder	0	⚪
gemma4 (gemma4)	embedding	0	⚪
gemma4 (gemma4)	vision_encoder	0	⚪
gemma4_text	model	0	⚪
gpt2	model	0	⚪
llama	model	0	⚪
llama (static-cache)	model	0	⚪
mamba (ssm-text-generation)	model	0	⚪
phi3	model	0	⚪
phi3 (static-cache)	model	0	⚪
qwen	model	0	⚪
qwen (static-cache)	model	0	⚪
qwen2	model	0	⚪
qwen2 (static-cache)	model	0	⚪
qwen2_moe	model	0	⚪
qwen2_moe (static-cache)	model	0	⚪
qwen3	model	0	⚪
qwen3 (static-cache)	model	0	⚪
qwen3_5_moe (hybrid-text-generation)	model	33	🔴
qwen3_5_text (hybrid-text-generation)	model	33	🔴
qwen3_5_vl (hybrid-qwen-vl)	decoder	56	🔴
qwen3_5_vl (hybrid-qwen-vl)	embedding	7	🟡
qwen3_5_vl (hybrid-qwen-vl)	vision_encoder	35	🟡
qwen3_moe	model	0	⚪
qwen3_moe (static-cache)	model	0	⚪
qwen3_next (hybrid-text-generation)	model	1	🔴
t5 (seq2seq)	decoder	0	⚪
t5 (seq2seq)	encoder	0	⚪
whisper (speech-to-text)	decoder	0	⚪
whisper (speech-to-text)	encoder	0	⚪

qwen3_5_moe (hybrid-text-generation) / model — 33 change(s)

Op summary: 269 → 265 nodes

--- base
+++ head
@@ -42,15 +42,11 @@
 Concat
 Reshape
 ReduceSumSquare
-CastLike
-Add
 Sqrt
 Div
 Reshape
 Reshape
 ReduceSumSquare
-CastLike
-Add
 Sqrt
 Div
 Reshape

Removed nodes:

• - CastLike
• - Add
• - CastLike
• - Add

Connectivity changes:

• node[4] Unsqueeze: input_ids [79, 11] → [78, 11]
• node[7] Sub: input_ids [82, 81] → [81, 80]
• node[8] Slice: input_ids [79, 83, 82, 11] → [78, 82, 81, 11]
• node[9] Unsqueeze: input_ids [84, 12] → [83, 12]
• node[10] GreaterOrEqual: input_ids [85, 80] → [84, 79]
• node[12] Cast: input_ids [87] → [86]
• node[13] And: input_ids [88, 86] → [87, 85]
• node[14] Where: input_ids [89, 13, 14] → [88, 13, 14]
• node[15] Cast: input_ids [90] → [89]
• node[16] Unsqueeze: input_ids [91, 11] → [90, 11]
• node[18] RMSNormalization: input_ids [76, 93] → [75, 92]
• node[19] Shape: input_ids [94] → [93]
• node[21] MatMul: input_ids [94, 96] → [93, 95]
• node[23] MatMul: input_ids [94, 98] → [93, 97]
• node[25] MatMul: input_ids [94, 100] → [93, 99]
• node[27] MatMul: input_ids [94, 102] → [93, 101]
• node[28] Transpose: input_ids [97] → [96]
• node[30] CastLike: input_ids [105, 23] → [104, 23]
• node[32] Expand: input_ids [106, 107] → [105, 106]
• node[33] CausalConvWithState: input_ids [104, 23, 108, 3] → [103, 23, 107, 3]
• node[34] Transpose: input_ids [109] → [108]
• node[36] Split: input_ids [111, 112] → [110, 111]
• node[37] Shape: input_ids [94] → [93]
• node[39] Concat: input_ids [95, 116, 117] → [94, 115, 116]
• node[41] Concat: input_ids [95, 116, 119] → [94, 115, 118]
• node[42] Reshape: input_ids [113, 118] → [112, 117]
• node[43] ReduceSumSquare: input_ids [121, 24] → [120, 24]

Interface changes:

• output[0]: dtype FLOAT → UNKNOWN; shape ['?', '?', '256'] → []; output[2]: dtype FLOAT → UNKNOWN; output[3]: dtype FLOAT → UNKNOWN; output[4]: dtype FLOAT → UNKNOWN

Initializer changes:

• count 69 → 68; dtype distribution: FLOAT: 62 → 61

qwen3_5_text (hybrid-text-generation) / model — 33 change(s)

Op summary: 131 → 127 nodes

--- base
+++ head
@@ -42,15 +42,11 @@
 Concat
 Reshape
 ReduceSumSquare
-CastLike
-Add
 Sqrt
 Div
 Reshape
 Reshape
 ReduceSumSquare
-CastLike
-Add
 Sqrt
 Div
 Reshape

Removed nodes:

• - CastLike
• - Add
• - CastLike
• - Add

Connectivity changes:

• node[4] Unsqueeze: input_ids [51, 11] → [50, 11]
• node[7] Sub: input_ids [54, 53] → [53, 52]
• node[8] Slice: input_ids [51, 55, 54, 11] → [50, 54, 53, 11]
• node[9] Unsqueeze: input_ids [56, 12] → [55, 12]
• node[10] GreaterOrEqual: input_ids [57, 52] → [56, 51]
• node[12] Cast: input_ids [59] → [58]
• node[13] And: input_ids [60, 58] → [59, 57]
• node[14] Where: input_ids [61, 13, 14] → [60, 13, 14]
• node[15] Cast: input_ids [62] → [61]
• node[16] Unsqueeze: input_ids [63, 11] → [62, 11]
• node[18] RMSNormalization: input_ids [48, 65] → [47, 64]
• node[19] Shape: input_ids [66] → [65]
• node[21] MatMul: input_ids [66, 68] → [65, 67]
• node[23] MatMul: input_ids [66, 70] → [65, 69]
• node[25] MatMul: input_ids [66, 72] → [65, 71]
• node[27] MatMul: input_ids [66, 74] → [65, 73]
• node[28] Transpose: input_ids [69] → [68]
• node[30] CastLike: input_ids [77, 23] → [76, 23]
• node[32] Expand: input_ids [78, 79] → [77, 78]
• node[33] CausalConvWithState: input_ids [76, 23, 80, 3] → [75, 23, 79, 3]
• node[34] Transpose: input_ids [81] → [80]
• node[36] Split: input_ids [83, 84] → [82, 83]
• node[37] Shape: input_ids [66] → [65]
• node[39] Concat: input_ids [67, 88, 89] → [66, 87, 88]
• node[41] Concat: input_ids [67, 88, 91] → [66, 87, 90]
• node[42] Reshape: input_ids [85, 90] → [84, 89]
• node[43] ReduceSumSquare: input_ids [93, 24] → [92, 24]

Interface changes:

• output[0]: dtype FLOAT → UNKNOWN; shape ['?', '?', '256'] → []; output[2]: dtype FLOAT → UNKNOWN; output[3]: dtype FLOAT → UNKNOWN; output[4]: dtype FLOAT → UNKNOWN

Initializer changes:

• count 41 → 40; dtype distribution: FLOAT: 34 → 33

qwen3_5_vl (hybrid-qwen-vl) / decoder — 56 change(s)

Op summary: 157 → 153 nodes

--- base
+++ head
@@ -67,15 +67,11 @@
 Concat
 Reshape
 ReduceSumSquare
-CastLike
-Add
 Sqrt
 Div
 Reshape
 Reshape
 ReduceSumSquare
-CastLike
-Add
 Sqrt
 Div
 Reshape

Removed nodes:

• - CastLike
• - Add
• - CastLike
• - Add

Connectivity changes:

• node[1] Reshape: input_ids [3, 53] → [3, 52]
• node[2] Transpose: input_ids [54] → [53]
• node[4] Gather: input_ids [55, 56] → [54, 55]
• node[6] Slice: input_ids [58, 12, 13, 14] → [57, 12, 13, 14]
• node[8] Concat: input_ids [60, 59] → [59, 58]
• node[10] Expand: input_ids [62, 61] → [61, 60]
• node[11] Reshape: input_ids [63, 61] → [62, 60]
• node[12] Gather: input_ids [64, 14] → [63, 14]
• node[13] Squeeze: input_ids [65, 14] → [64, 14]
• node[14] Gather: input_ids [8, 66] → [8, 65]
• node[15] Gather: input_ids [9, 66] → [9, 65]
• node[16] Gather: input_ids [64, 15] → [63, 15]
• node[17] Squeeze: input_ids [69, 14] → [68, 14]
• node[18] Gather: input_ids [8, 70] → [8, 69]
• node[19] Gather: input_ids [9, 70] → [9, 69]
• node[20] Gather: input_ids [64, 16] → [63, 16]
• node[21] Squeeze: input_ids [73, 14] → [72, 14]
• node[22] Gather: input_ids [8, 74] → [8, 73]
• node[23] Gather: input_ids [9, 74] → [9, 73]
• node[24] Where: input_ids [10, 71, 67] → [10, 70, 66]
• node[25] Where: input_ids [11, 75, 77] → [11, 74, 76]
• node[26] Where: input_ids [10, 72, 68] → [10, 71, 67]
• node[27] Where: input_ids [11, 76, 79] → [11, 75, 78]
• node[29] Unsqueeze: input_ids [81, 15] → [80, 15]
• node[32] Sub: input_ids [84, 83] → [83, 82]
• node[33] Slice: input_ids [81, 85, 84, 15] → [80, 84, 83, 15]
• node[34] Unsqueeze: input_ids [86, 16] → [85, 16]
• node[35] GreaterOrEqual: input_ids [87, 82] → [86, 81]
• node[37] Cast: input_ids [89] → [88]
• node[38] And: input_ids [90, 88] → [89, 87]
• node[39] Where: input_ids [91, 18, 19] → [90, 18, 19]
• node[40] Cast: input_ids [92] → [91]
• node[41] Unsqueeze: input_ids [93, 15] → [92, 15]
• node[43] RMSNormalization: input_ids [0, 95] → [0, 94]
• node[44] Shape: input_ids [96] → [95]
• node[46] MatMul: input_ids [96, 98] → [95, 97]
• node[48] MatMul: input_ids [96, 100] → [95, 99]
• node[50] MatMul: input_ids [96, 102] → [95, 101]
• node[52] MatMul: input_ids [96, 104] → [95, 103]
• node[53] Transpose: input_ids [99] → [98]
• node[55] CastLike: input_ids [107, 28] → [106, 28]
• node[57] Expand: input_ids [108, 109] → [107, 108]
• node[58] CausalConvWithState: input_ids [106, 28, 110, 4] → [105, 28, 109, 4]
• node[59] Transpose: input_ids [111] → [110]
• node[61] Split: input_ids [113, 114] → [112, 113]
• node[62] Shape: input_ids [96] → [95]
• node[64] Concat: input_ids [97, 118, 119] → [96, 117, 118]
• node[66] Concat: input_ids [97, 118, 121] → [96, 117, 120]
• node[67] Reshape: input_ids [115, 120] → [114, 119]
• node[68] ReduceSumSquare: input_ids [123, 29] → [122, 29]

Interface changes:

• output[0]: dtype FLOAT → UNKNOWN; shape ['?', '?', '256'] → []; output[2]: dtype FLOAT → UNKNOWN; output[3]: dtype FLOAT → UNKNOWN; output[4]: dtype FLOAT → UNKNOWN

Initializer changes:

• count 45 → 44; dtype distribution: FLOAT: 33 → 32

qwen3_5_vl (hybrid-qwen-vl) / embedding — 7 change(s)

Op summary: 52 → 46 nodes

--- base
+++ head
@@ -12,20 +12,14 @@
 Gather
 Constant
 Equal
-Constant
-Equal
-Or
 Unsqueeze
 Cast
-Reshape
 Constant
 CumSum
 Constant
 Sub
 Constant
 Clip
-Shape
-Reshape
 CastLike
 Constant
 Shape

Removed nodes:

• - Constant
• - Equal
• - Or
• - Reshape
• - Shape
• - Reshape

Connectivity changes:

• node[42] Concat: input_ids [46, 47] → [47]

qwen3_5_vl (hybrid-qwen-vl) / vision_encoder — 35 change(s)

Op summary: 252 → 251 nodes

--- base
+++ head
@@ -1,4 +1,3 @@
-Cast
 Reshape
 Conv
 Reshape

Removed nodes:

• - Cast

Modified attributes:

• node[93] Constant: value_int: None → 0, value_ints: [1, 0] → None
• node[127] Constant: value_int: 0 → 1
• node[202] Constant: value_int: 0 → 1

Connectivity changes:

• node[10] Mul: input_ids [64, 66] → [61, 66]
• node[21] Unsqueeze: input_ids [77, 7] → [67, 8]
• node[39] Mul: input_ids [93, 95] → [90, 95]
• node[50] Unsqueeze: input_ids [106, 7] → [96, 8]
• node[60] Gather: input_ids [116, 7] → [115, 8]
• node[64] Gather: input_ids [12, 119] → [13, 118]
• node[65] Gather: input_ids [13, 119] → [12, 119]
• node[66] Gather: input_ids [12, 120] → [13, 119]
• node[81] Unsqueeze: input_ids [137, 7] → [127, 8]
• node[103] CastLike: input_ids [125, 160] → [125, 159]
• node[112] Mul: input_ids [165, 166] → [163, 166]
• node[119] Mul: input_ids [165, 168] → [163, 168]
• node[123] Reshape: input_ids [176, 20] → [182, 20]
• node[130] Unsqueeze: input_ids [190, 8] → [151, 7]
• node[137] Unsqueeze: input_ids [197, 8] → [196, 7]
• node[140] CastLike: input_ids [21, 184] → [22, 183]
• node[143] Unsqueeze: input_ids [203, 23] → [183, 7]
• node[145] Unsqueeze: input_ids [185, 7] → [158, 7]
• node[151] Add: input_ids [211, 25] → [88, 211]
• node[160] Add: input_ids [220, 31] → [212, 220]
• node[178] CastLike: input_ids [125, 239] → [125, 238]
• node[187] Mul: input_ids [244, 245] → [242, 245]
• node[194] Mul: input_ids [244, 247] → [242, 247]
• node[198] Reshape: input_ids [255, 20] → [261, 20]
• node[205] Unsqueeze: input_ids [269, 8] → [151, 7]
• node[212] Unsqueeze: input_ids [276, 8] → [275, 7]
• node[215] CastLike: input_ids [21, 263] → [22, 262]
• node[218] Unsqueeze: input_ids [282, 23] → [262, 7]
• node[220] Unsqueeze: input_ids [264, 7] → [237, 7]
• node[226] Add: input_ids [290, 44] → [221, 290]
• node[235] Add: input_ids [299, 50] → [291, 299]

qwen3_next (hybrid-text-generation) / model — 1 change(s)

Op summary: 565 → 565 nodes

No op-sequence changes.

Interface changes:

• output[0]: dtype FLOAT → UNKNOWN; shape ['?', '?', '256'] → []; output[2]: dtype FLOAT → UNKNOWN; output[4]: dtype FLOAT → UNKNOWN; output[6]: dtype FLOAT → UNKNOWN; output[7]: dtype FLOAT → UNKNOWN; output[8]: dtype FLOAT → UNKNOWN

---

Legend: ⚪ No change · 🔵 Minor (attrs/inits) · 🟡 Moderate (nodes added/removed) · 🔴 Major (interface changed)

[justinchuby]

LGTM. Could you add the motivation in the PR description? Also cc @titaiwangms for another review.

[titaiwangms]

Full review — changes requested

Reviewed commit: 06e009786988f5455575b10a24e02f38e72d7175. Five independent static reviewers plus focused local verification. This review concerns that revision only.

Confirmed findings

F2 — Major: global per-channel GPTQ loading loses logical K
src/mobius/_weight_utils.py:896–906

The global quantization path does not pass in_features or defer normalization to the leaf codec. For a GPTQ down projection with K=33, bits=4, and group_size=-1, five int32 words are interpreted as K=40. Preprocessing produces [64, 1, 20], while the module expects [64, 1, 17]. Reproduced through CausalLMModel.preprocess_weights(), build_from_module(), and apply_weights(), which raises ValueError: Weight shape mismatch. Normalize against each target linear's logical K on the global path as well; do not infer logical K from padded storage.

F3 — Major: explicit group_size=K is accepted but cannot load the same layout as -1
src/mobius/components/_quantized_linear.py:133–148; src/mobius/_weight_utils.py:1494–1508

For K=33, bits=4, QuantizedLinear accepts an explicit full-channel block and allocates 17 bytes. Olive normalization only special-cases group_size=-1; explicit 33 uses floor division and a 16-byte blob, rejecting the valid 17-byte row. For K=group_size=1, it raises ZeroDivisionError. Both failures were reproduced; the equivalent -1 cases succeed. Treat explicit full-channel groups and -1 consistently using padding-aware normalization, or reject the explicit form consistently.

F4 — Major, with reviewer dissent: selected CPU export retains unsupported native block layouts
src/mobius/_optimizations.py:571–576, enabled by src/mobius/components/_quantized_linear.py:133–148

On ONNX Runtime 1.30.0 CPU, native MatMulNBits rejects K=block_size values 7, 96, and 1536. The corresponding explicitly inlined cases all pass. The optimizer checks only the backend-wide supports_matmul_nbits, so CPU/default exports retain the native op without checking layout compatibility. Preserve arbitrary-K graph construction for compatible targets, but lower unsupported layouts or clearly reject incompatible EP/layout combinations before export. CUDA and ORT 1.31 native behavior were not verified.

Dissent: the Critical reviewer ultimately considers this a nonblocking backend-policy question because the constructor documents backend-specific support and the export compatibility promise is not explicit. The Deep reviewer and review lead retain Major specifically for the selected-CPU export path, not because arbitrary-K representation is intrinsically invalid.

F1 — Minor: the new K=1536 parity test fails locally and in CI
src/mobius/functions/matmul_nbits_test.py:206–234

test_inline_trims_single_block_padding[False-1536-1536] fails for 1 of 48 outputs at rtol=atol=1e-5: expected 0.4662437438964844, actual 0.4662666320800781, maximum absolute difference 2.2888184e-05. Reproduced locally on ORT 1.30.0 and observed in the pinned-head Ubuntu CI run 40462101700 on ORT 1.31.1. This establishes a failing test, not weight corruption or a proven numerical implementation defect. Isolate dequantization versus accumulation differences before deciding the appropriate fix; do not simply relax the tolerance without diagnosis.

Findings ledger

ID	Severity	Raised by	Disposition
F2	Major	Code / Critical / Integration	Confirmed by public build/apply-weights reproduction
F3	Major	Deep / Integration	Confirmed by K=33 and K=1 normalization reproductions
F4	Major, disputed	Critical initially / Deep	Confirmed CPU compatibility gap; Critical's later policy-severity dissent preserved above
F1	Minor	CI / local verification	Confirmed test failure; numerical root cause not established

Verification and remaining questions

Focused local tests: 687 passed, 1 failed, 2 skipped. Additional reproductions covered global logical-K loading, explicit full-channel loading, and native-versus-inline execution. No independent accuracy-override propagation defect was found in the updated Python builders.

Nonblocking questions:

• Deep / Critical: producer-derived asymmetric GPTQ/AWQ fixtures with distinct channel codes and zero points are still needed to establish real-producer compatibility; no corruption claim is established.
• Integration: should CLI/auto_export expose the new accuracy override, and is per-channel QMoE intentionally unsupported? These are scope/policy questions, not confirmed regressions.

Not checked: CUDA/OpenVINO execution, real checkpoints, the full test suite, or a full security audit. The Windows GGUF CI failure was not established as caused by this PR. No implementation changes were made.