hf2q 0.1.17

Pure Rust CLI for converting HuggingFace models to hardware-optimized formats and serving them over an OpenAI-compatible API on Apple Silicon
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
use anyhow::{bail, ensure, Context, Result};
use mlx_native::gguf::GgufFile;
use mlx_native::{MlxBuffer, MlxDevice};

use super::ffn::{DenseFfnWeights, MoeFfnShape};
use super::gpu_ffn::{DenseFfnWeightsGpu, DenseFfnWeightsGpuQ, MoeFfnWeightsGpuQ};
use super::gpu_full_attn::{upload_f32_weight, FullAttnQGateWeightsGpu};
use super::mtp::{MtpFfnWeightsGpu, MtpFullAttnWeightsGpu, MtpQGateWeightsGpu, MtpWeights};
use super::weight_loader::{
    dense_ffn_storage, dense_ffn_tensor_types, load_dense_ffn_quantized, load_f32_tensor,
    load_moe_ffn_quantized, load_native_projection, qwen35_moe_expert_type_supported,
    validate_native_projection_info, DenseFfnStorage,
};
use super::Qwen35Config;
use crate::serve::forward_mlx_shared::MlxQWeight;

pub fn load_mtp_weights_if_present(
    gguf: &GgufFile,
    cfg: &Qwen35Config,
    device: &MlxDevice,
) -> Result<Option<MtpWeights>> {
    load_mtp_weights_if_present_with_shared_head(gguf, cfg, device, None)
}

pub fn load_mtp_weights_if_present_with_shared_head(
    gguf: &GgufFile,
    cfg: &Qwen35Config,
    device: &MlxDevice,
    main_output_head: Option<&MlxQWeight>,
) -> Result<Option<MtpWeights>> {
    if cfg.mtp_num_hidden_layers == 0 {
        return Ok(None);
    }
    validate_mtp_tensor_topology(gguf, cfg)?;

    let layer_index = cfg.num_hidden_layers;
    let loaded_tensor_names = mtp_tensor_names(gguf, layer_index);

    let h = cfg.hidden_size as usize;
    let p = format!("blk.{layer_index}");
    let nextn = format!("{p}.nextn");
    let enorm = load_norm_gpu(gguf, &format!("{nextn}.enorm.weight"), h, device)?;
    let hnorm = load_norm_gpu(gguf, &format!("{nextn}.hnorm.weight"), h, device)?;
    // ADR-013 P14 follow-up (2026-04-30): honor mtp_use_dedicated_embeddings.
    //
    // - True  → MTP carries its own embed table at `blk.{N}.nextn.embed_tokens.weight`
    //           (Qwen3.5 MTP convention).
    // - False → MTP shares the main model's `token_embd.weight` (Qwen3.6 27B + 35B-A3B
    //           convention; convert correctly skips emitting the redundant tensor).
    //           We do NOT duplicate the buffer: `forward_draft` is called with the
    //           per-token embedding already materialised by the verifier's hot path,
    //           and the field itself is reserved for future direct lookups (use the
    //           main model's token_embd via Qwen35Model accessors).
    //
    // Logged at INFO level so operators can confirm the path.
    let embed_tokens_tname = format!("{nextn}.embed_tokens.weight");
    let embed_tokens = if cfg.mtp_use_dedicated_embeddings {
        let info = gguf.tensor_info(&embed_tokens_tname).ok_or_else(|| {
            anyhow::anyhow!("MTP dedicated embedding tensor `{embed_tokens_tname}` is missing")
        })?;
        ensure!(
            info.shape.len() == 2 && info.shape[1] == h,
            "{embed_tokens_tname} shape {:?} is not [vocab,{h}]",
            info.shape
        );
        let buf =
            crate::serve::forward_mlx_shared::load_gguf_qweight(gguf, &embed_tokens_tname, device)
                .with_context(|| {
                    format!(
                        "MTP loader expected dedicated `{embed_tokens_tname}` because \
                 mtp_use_dedicated_embeddings=True"
                    )
                })?;
        super::forward_gpu::ensure_native_embedding_admitted(&buf)
            .with_context(|| format!("admit direct execution for {embed_tokens_tname}"))?;
        super::weight_pool::register_weight_buffer(device, &buf.buffer)
            .with_context(|| format!("register {embed_tokens_tname}"))?;
        tracing::info!(
            mtp_layer = layer_index,
            mtp_use_dedicated_embeddings = true,
            tensor = %embed_tokens_tname,
            "qwen35 MTP loader: dedicated embed_tokens"
        );
        Some(buf)
    } else {
        // Belt-and-suspenders: if convert-side regression ever re-emits the dedicated
        // tensor while the flag says shared, refuse to silently ignore one of them.
        if gguf.tensor_info(&embed_tokens_tname).is_some() {
            bail!(
                "qwen35 MTP loader: mtp_use_dedicated_embeddings=False but `{embed_tokens_tname}` \
                 is present in the GGUF — convert pipeline is inconsistent. Re-emit without \
                 the dedicated tensor or set the metadata key to true."
            );
        }
        tracing::info!(
            mtp_layer = layer_index,
            mtp_use_dedicated_embeddings = false,
            "qwen35 MTP loader: sharing main token_embd (no dedicated nextn.embed_tokens)"
        );
        None
    };
    let (eh_proj, eh_proj_ggml_type) =
        load_native_projection(gguf, &format!("{nextn}.eh_proj.weight"), h, 2 * h, device)?;
    let shared_head_norm =
        load_norm_gpu(gguf, &format!("{nextn}.shared_head_norm.weight"), h, device)?;

    // ADR-013 P14 follow-up (2026-04-30): the LM-head projection weight (`shared_head.head`)
    // follows the same shared-vs-dedicated rule as `embed_tokens`.
    //
    // Qwen3.6 27B + 35B-A3B (`mtp_use_dedicated_embeddings: False`) ship neither
    // `mtp.embed_tokens` nor `mtp.shared_head.head`; the MTP block reuses the main
    // verifier's `token_embd.weight` for the embedding lookup and resolves the
    // main output projection from `output.weight` when present, otherwise from
    // that same tied token-embedding allocation. Convert correctly skips emitting
    // `blk.{N}.nextn.shared_head_head.weight` in this configuration.
    //
    // Resolution: if the dedicated tensor is present we use its native GGUF
    // buffer (Qwen3.5 MTP); otherwise shared mode borrows the resolved main
    // head allocation. `vocab_size` is derived from the row count of the exact
    // tensor selected by that rule.
    let shared_head_head_tname = format!("{nextn}.shared_head_head.weight");
    let (shared_head_head, shared_head_head_ggml_type, vocab_size, shared_head_head_source) =
        if let Some(info) = gguf.tensor_info(&shared_head_head_tname) {
            ensure!(
                info.shape.len() == 2 && info.shape[1] == h,
                "{shared_head_head_tname} shape {:?} is not [vocab, {h}]",
                info.shape
            );
            let vocab = info.shape[0];
            let (buffer, ggml_type) =
                load_native_projection(gguf, &shared_head_head_tname, vocab, h, device)?;
            (
                buffer,
                ggml_type,
                vocab as u32,
                shared_head_head_tname.clone(),
            )
        } else if !cfg.mtp_use_dedicated_embeddings {
            // Shared mode: borrow the physical main LM head. GGUF represents
            // tied Qwen3.5 heads by omitting output.weight; in that case the
            // exact token_embd.weight blocks are the authoritative head.
            let main_lm = if gguf.tensor_info("output.weight").is_some() {
                "output.weight"
            } else {
                "token_embd.weight"
            };
            if let Some(main) = main_output_head {
                ensure!(
                    main.affine.is_none() && main.info.cols == h,
                    "qwen35 MTP loader: supplied shared output head is not a native [vocab,{h}] GGUF projection"
                );
                (
                    main.buffer.clone(),
                    main.info.ggml_dtype,
                    main.info.rows as u32,
                    main_lm.to_string(),
                )
            } else {
                let info = gguf.tensor_info(main_lm).ok_or_else(|| {
                    anyhow::anyhow!(
                        "qwen35 MTP loader: shared head missing and resolved main head {main_lm} absent"
                    )
                })?;
                ensure!(
                    info.shape.len() == 2 && info.shape[1] == h,
                    "{main_lm} shape {:?} is not [vocab, {h}]",
                    info.shape
                );
                let vocab = info.shape[0];
                let (buffer, ggml_type) = load_native_projection(gguf, main_lm, vocab, h, device)?;
                (buffer, ggml_type, vocab as u32, main_lm.to_string())
            }
        } else {
            bail!(
                "qwen35 MTP loader: `{shared_head_head_tname}` is missing AND \
                 mtp_use_dedicated_embeddings=True — cannot resolve the MTP final \
                 projection. Re-emit the GGUF with the dedicated tensor or set the \
                 flag to false."
            );
        };
    tracing::info!(
        mtp_layer = layer_index,
        source = %shared_head_head_source,
        vocab_size,
        "qwen35 MTP loader: shared_head_head resolved"
    );
    let attn = load_mtp_attn(gguf, cfg, layer_index, device)?;
    let (ffn, intermediate_size) = load_mtp_ffn(gguf, cfg, layer_index, device)?;

    Ok(Some(MtpWeights {
        layer_index,
        hidden_size: cfg.hidden_size,
        vocab_size,
        intermediate_size,
        loaded_tensor_names,
        enorm,
        hnorm,
        eh_proj,
        eh_proj_ggml_type,
        embed_tokens,
        shared_head_norm,
        shared_head_head,
        shared_head_head_ggml_type,
        attn,
        ffn,
    }))
}

/// Descriptor-only MTP contract shared by bounded hosted admission and the
/// production loader. It validates every required tensor name/shape and the
/// dense-vs-MoE exclusivity/storage relation without touching tensor payloads.
pub(crate) fn validate_mtp_tensor_topology(gguf: &GgufFile, cfg: &Qwen35Config) -> Result<()> {
    if cfg.mtp_num_hidden_layers == 0 {
        return Ok(());
    }
    ensure!(
        cfg.mtp_num_hidden_layers == 1,
        "qwen35 MTP loader supports exactly one nextn layer, got {}",
        cfg.mtp_num_hidden_layers
    );
    let layer = cfg.num_hidden_layers;
    ensure!(
        !mtp_tensor_names(gguf, layer).is_empty(),
        "qwen35 metadata advertises nextn_predict_layers=1 but no blk.{layer}.nextn.* or blk.{layer}.* MTP tensors were found"
    );
    let h = cfg.hidden_size as usize;
    let q = (cfg.num_attention_heads as usize)
        .checked_mul(cfg.head_dim as usize)
        .context("MTP Q projection dimension overflow")?;
    let kv = (cfg.num_key_value_heads as usize)
        .checked_mul(cfg.head_dim as usize)
        .context("MTP KV projection dimension overflow")?;
    let d = cfg.head_dim as usize;
    let p = format!("blk.{layer}");
    let nextn = format!("{p}.nextn");
    let require_shape = |name: &str, expected: &[usize]| -> Result<()> {
        let info = gguf
            .tensor_info(name)
            .ok_or_else(|| anyhow::anyhow!("required MTP tensor `{name}` is missing"))?;
        ensure!(
            info.shape.as_slice() == expected,
            "MTP tensor `{name}` shape {:?} != {expected:?}",
            info.shape
        );
        Ok(())
    };
    for name in [
        format!("{nextn}.enorm.weight"),
        format!("{nextn}.hnorm.weight"),
        format!("{nextn}.shared_head_norm.weight"),
        format!("{p}.attn_norm.weight"),
        format!("{p}.post_attention_norm.weight"),
    ] {
        require_shape(&name, &[h])?;
    }
    let eh_name = format!("{nextn}.eh_proj.weight");
    let eh_info = gguf
        .tensor_info(&eh_name)
        .ok_or_else(|| anyhow::anyhow!("required MTP tensor `{eh_name}` is missing"))?;
    validate_native_projection_info(&eh_name, eh_info, h, 2 * h)?;

    let embed = format!("{nextn}.embed_tokens.weight");
    let embedding_rows = match (cfg.mtp_use_dedicated_embeddings, gguf.tensor_info(&embed)) {
        (true, Some(info)) => {
            ensure!(
                info.shape.len() == 2 && info.shape[1] == h,
                "MTP dedicated embedding tensor `{embed}` shape {:?} is not [vocab,{h}]",
                info.shape
            );
            super::forward_gpu::validate_native_embedding_descriptor(
                &embed,
                info.ggml_type,
                info.shape[0],
                h,
                info.byte_len,
            )?;
            info.shape[0]
        }
        (true, None) => bail!("MTP dedicated embedding tensor `{embed}` is missing"),
        (false, Some(_)) => {
            bail!("qwen35 MTP loader: mtp_use_dedicated_embeddings=False but `{embed}` is present")
        }
        (false, None) => {
            gguf.tensor_info("token_embd.weight")
                .context("MTP shared embedding tensor `token_embd.weight` is missing")?
                .shape[0]
        }
    };
    let head = format!("{nextn}.shared_head_head.weight");
    let head_rows = if let Some(info) = gguf.tensor_info(&head) {
        ensure!(
            info.shape.len() == 2,
            "MTP shared head tensor `{head}` is not a matrix"
        );
        validate_native_projection_info(&head, info, info.shape[0], h)?;
        info.shape[0]
    } else {
        ensure!(
            !cfg.mtp_use_dedicated_embeddings,
            "qwen35 MTP loader: `{head}` is missing while dedicated embeddings are enabled"
        );
        let main_head = if gguf.tensor_info("output.weight").is_some() {
            "output.weight"
        } else {
            "token_embd.weight"
        };
        let info = gguf
            .tensor_info(main_head)
            .ok_or_else(|| anyhow::anyhow!("MTP shared main head `{main_head}` is missing"))?;
        ensure!(
            info.shape.len() == 2,
            "MTP shared main head `{main_head}` is not a matrix"
        );
        validate_native_projection_info(main_head, info, info.shape[0], h)?;
        info.shape[0]
    };
    ensure!(
        embedding_rows >= head_rows,
        "MTP embedding rows {embedding_rows} cannot cover selected MTP head rows {head_rows}"
    );

    for (name, rows, cols) in [
        (format!("{p}.attn_k.weight"), kv, h),
        (format!("{p}.attn_v.weight"), kv, h),
        (format!("{p}.attn_output.weight"), h, q),
    ] {
        let info = gguf
            .tensor_info(&name)
            .ok_or_else(|| anyhow::anyhow!("required MTP tensor `{name}` is missing"))?;
        validate_native_projection_info(&name, info, rows, cols)?;
    }
    require_shape(&format!("{p}.attn_q_norm.weight"), &[d])?;
    require_shape(&format!("{p}.attn_k_norm.weight"), &[d])?;
    let q_name = format!("{p}.attn_q.weight");
    let q_info = gguf
        .tensor_info(&q_name)
        .ok_or_else(|| anyhow::anyhow!("required MTP tensor `{q_name}` is missing"))?;
    ensure!(
        q_info.shape.as_slice() == [q, h] || q_info.shape.as_slice() == [2 * q, h],
        "MTP tensor `{q_name}` shape {:?} is neither [{q},{h}] nor [{},{h}]",
        q_info.shape,
        2 * q
    );
    validate_native_projection_info(&q_name, q_info, q_info.shape[0], h)?;
    if q_info.shape.as_slice() == [q, h] {
        let gate_name = format!("{p}.attn_gate.weight");
        if let Some(gate_info) = gguf.tensor_info(&gate_name) {
            validate_native_projection_info(&gate_name, gate_info, q, h)?;
        }
    } else {
        ensure!(
            gguf.tensor_info(&format!("{p}.attn_gate.weight")).is_none(),
            "MTP fused Q+gate tensor cannot also have a split attn_gate tensor"
        );
    }

    let has_dense = gguf.tensor_info(&format!("{p}.ffn_gate.weight")).is_some();
    let has_moe = gguf
        .tensor_info(&format!("{p}.ffn_gate_exps.weight"))
        .is_some();
    ensure!(
        has_dense ^ has_moe,
        "qwen35 MTP block {layer} must have exactly one dense or MoE FFN layout"
    );
    if has_dense {
        let gate = gguf
            .tensor_info(&format!("{p}.ffn_gate.weight"))
            .expect("presence checked above");
        ensure!(
            gate.shape.len() == 2 && gate.shape[1] == h,
            "MTP dense gate shape {:?} is not [intermediate,{h}]",
            gate.shape
        );
        let intermediate = gate.shape[0];
        require_shape(&format!("{p}.ffn_up.weight"), &[intermediate, h])?;
        require_shape(&format!("{p}.ffn_down.weight"), &[h, intermediate])?;
        let (gate_type, up_type, down_type) = dense_ffn_tensor_types(gguf, layer)?;
        dense_ffn_storage(layer, gate_type, up_type, down_type)?;
    } else {
        let moe = cfg
            .moe
            .as_ref()
            .context("MTP MoE tensors require MoE runtime metadata")?;
        let experts = moe.num_experts as usize;
        let intermediate = moe.moe_intermediate_size as usize;
        let shared = moe.shared_expert_intermediate_size as usize;
        require_shape(&format!("{p}.ffn_gate_inp.weight"), &[experts, h])?;
        require_shape(
            &format!("{p}.ffn_gate_exps.weight"),
            &[experts, intermediate, h],
        )?;
        require_shape(
            &format!("{p}.ffn_up_exps.weight"),
            &[experts, intermediate, h],
        )?;
        require_shape(
            &format!("{p}.ffn_down_exps.weight"),
            &[experts, h, intermediate],
        )?;
        require_shape(&format!("{p}.ffn_gate_inp_shexp.weight"), &[1, h])?;
        require_shape(&format!("{p}.ffn_gate_shexp.weight"), &[shared, h])?;
        require_shape(&format!("{p}.ffn_up_shexp.weight"), &[shared, h])?;
        require_shape(&format!("{p}.ffn_down_shexp.weight"), &[h, shared])?;
        let gate_name = format!("{p}.ffn_gate_exps.weight");
        let up_name = format!("{p}.ffn_up_exps.weight");
        let down_name = format!("{p}.ffn_down_exps.weight");
        let gate = gguf
            .tensor_info(&gate_name)
            .expect("MTP MoE gate presence checked above");
        let up = gguf
            .tensor_info(&up_name)
            .expect("MTP MoE up shape checked above");
        let down = gguf
            .tensor_info(&down_name)
            .expect("MTP MoE down shape checked above");
        ensure!(
            gate.ggml_type == up.ggml_type,
            "MTP MoE gate/up quant types differ ({:?} vs {:?}); both buffers share one dispatch type",
            gate.ggml_type,
            up.ggml_type
        );
        ensure!(
            qwen35_moe_expert_type_supported(gate.ggml_type),
            "MTP MoE gate/up expert weights use unsupported {:?} storage",
            gate.ggml_type
        );
        ensure!(
            qwen35_moe_expert_type_supported(down.ggml_type),
            "MTP MoE down expert weights use unsupported {:?} storage",
            down.ggml_type
        );
    }
    Ok(())
}

pub(super) fn mtp_tensor_names(gguf: &GgufFile, layer_index: u32) -> Vec<String> {
    let p = format!("blk.{layer_index}.");
    let nextn = format!("{p}nextn.");
    // Inner-block tensor suffixes for both dense MTP (Qwen 3.6 27B) and MoE
    // MTP (Qwen 3.5/3.6 35B-A3B). The two FFN schemas are mutually exclusive
    // at a given block, so listing both in the membership set is safe.
    let inner = [
        // Attention (shared by both variants).
        "attn_norm.weight",
        "post_attention_norm.weight",
        "attn_q.weight",
        "attn_gate.weight",
        "attn_k.weight",
        "attn_v.weight",
        "attn_output.weight",
        "attn_q_norm.weight",
        "attn_k_norm.weight",
        // Dense FFN (Qwen 3.6 27B dense-MTP).
        "ffn_gate.weight",
        "ffn_up.weight",
        "ffn_down.weight",
        // MoE FFN (Qwen 3.5/3.6 35B-A3B MoE-MTP).
        "ffn_gate_inp.weight",
        "ffn_gate_exps.weight",
        "ffn_up_exps.weight",
        "ffn_down_exps.weight",
        "ffn_gate_inp_shexp.weight",
        "ffn_gate_shexp.weight",
        "ffn_up_shexp.weight",
        "ffn_down_shexp.weight",
    ];
    let mut names = Vec::new();
    for name in gguf.tensor_names() {
        if name.starts_with(&nextn) || inner.iter().any(|suffix| name == format!("{p}{suffix}")) {
            names.push(name.to_string());
        }
    }
    names.sort();
    names
}

fn load_mtp_attn(
    gguf: &GgufFile,
    cfg: &Qwen35Config,
    layer_index: u32,
    device: &MlxDevice,
) -> Result<MtpFullAttnWeightsGpu> {
    let p = format!("blk.{layer_index}");
    let h = cfg.hidden_size as usize;
    let q_total = (cfg.num_attention_heads * cfg.head_dim) as usize;
    let kv_total = (cfg.num_key_value_heads * cfg.head_dim) as usize;
    let d = cfg.head_dim as usize;
    let attn_norm = load_norm_gpu(gguf, &format!("{p}.attn_norm.weight"), h, device)?;
    let post_attn_norm =
        load_norm_gpu(gguf, &format!("{p}.post_attention_norm.weight"), h, device)?;

    let q_name = format!("{p}.attn_q.weight");
    let q_info = gguf
        .tensor_info(&q_name)
        .ok_or_else(|| anyhow::anyhow!("{q_name} not found"))?;
    let q_gate = if q_info.shape.as_slice() == [q_total, h] {
        let (q, q_type) = load_native_projection(gguf, &q_name, q_total, h, device)?;
        let gate_name = format!("{p}.attn_gate.weight");
        if gguf.tensor_info(&gate_name).is_some() {
            let (gate, gate_type) = load_native_projection(gguf, &gate_name, q_total, h, device)?;
            MtpQGateWeightsGpu::Gated(FullAttnQGateWeightsGpu::Split {
                wq: q,
                wq_ggml_type: q_type,
                w_gate: gate,
                w_gate_ggml_type: gate_type,
            })
        } else {
            MtpQGateWeightsGpu::Ungated {
                wq: q,
                wq_ggml_type: q_type,
            }
        }
    } else if q_info.shape.as_slice() == [2 * q_total, h] {
        let (fused, q_type) = load_native_projection(gguf, &q_name, 2 * q_total, h, device)?;
        MtpQGateWeightsGpu::Gated(FullAttnQGateWeightsGpu::Fused {
            weight: fused,
            ggml_type: q_type,
        })
    } else {
        bail!(
            "{q_name} shape {:?}, expected [{q_total},{h}] or [{}, {h}] interleaved Q+gate",
            q_info.shape,
            2 * q_total,
        );
    };

    let (wk, wk_ggml_type) =
        load_native_projection(gguf, &format!("{p}.attn_k.weight"), kv_total, h, device)?;
    let (wv, wv_ggml_type) =
        load_native_projection(gguf, &format!("{p}.attn_v.weight"), kv_total, h, device)?;
    let (wo, wo_ggml_type) =
        load_native_projection(gguf, &format!("{p}.attn_output.weight"), h, q_total, device)?;

    Ok(MtpFullAttnWeightsGpu {
        attn_norm,
        post_attn_norm,
        q_gate,
        wk,
        wk_ggml_type,
        wv,
        wv_ggml_type,
        attn_q_norm: load_norm_gpu(gguf, &format!("{p}.attn_q_norm.weight"), d, device)?,
        attn_k_norm: load_norm_gpu(gguf, &format!("{p}.attn_k_norm.weight"), d, device)?,
        wo,
        wo_ggml_type,
    })
}

fn load_mtp_ffn(
    gguf: &GgufFile,
    cfg: &Qwen35Config,
    layer_index: u32,
    device: &MlxDevice,
) -> Result<(MtpFfnWeightsGpu, u32)> {
    let p = format!("blk.{layer_index}");
    let has_dense = gguf.tensor_info(&format!("{p}.ffn_gate.weight")).is_some();
    let has_moe = gguf
        .tensor_info(&format!("{p}.ffn_gate_exps.weight"))
        .is_some();
    match (has_dense, has_moe) {
        (true, false) => load_mtp_dense_ffn(gguf, cfg, &p, device),
        (false, true) => load_mtp_moe_ffn(gguf, cfg, layer_index, &p, device),
        (true, true) => bail!(
            "qwen35 MTP loader: block {layer_index} has BOTH dense (`{p}.ffn_gate.weight`) and \
             MoE (`{p}.ffn_gate_exps.weight`) FFN tensors — GGUF is malformed"
        ),
        (false, false) => bail!(
            "qwen35 MTP loader: block {layer_index} has NEITHER dense (`{p}.ffn_gate.weight`) \
             nor MoE (`{p}.ffn_gate_exps.weight`) inner FFN tensors"
        ),
    }
}

fn load_mtp_dense_ffn(
    gguf: &GgufFile,
    cfg: &Qwen35Config,
    p: &str,
    device: &MlxDevice,
) -> Result<(MtpFfnWeightsGpu, u32)> {
    let layer_idx = cfg.num_hidden_layers;
    let (gate_type, up_type, down_type) = dense_ffn_tensor_types(gguf, layer_idx)?;
    match dense_ffn_storage(layer_idx, gate_type, up_type, down_type)? {
        DenseFfnStorage::Quantized => {
            let weights_q = load_dense_ffn_quantized(gguf, layer_idx, cfg, device)
                .with_context(|| format!("MTP native dense FFN {p}"))?;
            let intermediate_size = weights_q.intermediate_size;
            let dense_gpu = DenseFfnWeightsGpuQ::from_quantized(&weights_q);
            Ok((
                MtpFfnWeightsGpu::DenseQ { weights: dense_gpu },
                intermediate_size,
            ))
        }
        DenseFfnStorage::Float => {
            let h = cfg.hidden_size as usize;
            let gate_name = format!("{p}.ffn_gate.weight");
            let gate = load_f32_tensor(gguf, &gate_name, device)?;
            ensure!(gate.len() % h == 0, "{gate_name} width mismatch");
            let intermediate = gate.len() / h;
            let up = load_f32_tensor(gguf, &format!("{p}.ffn_up.weight"), device)?;
            let down = load_f32_tensor(gguf, &format!("{p}.ffn_down.weight"), device)?;
            ensure!(
                up.len() == intermediate * h,
                "{p}.ffn_up.weight shape mismatch"
            );
            ensure!(
                down.len() == h * intermediate,
                "{p}.ffn_down.weight shape mismatch"
            );
            let weights =
                DenseFfnWeightsGpu::from_cpu(&DenseFfnWeights { gate, up, down }, device)?;
            let intermediate_size = intermediate as u32;
            Ok((
                MtpFfnWeightsGpu::Dense {
                    weights,
                    intermediate_size,
                },
                intermediate_size,
            ))
        }
    }
}

fn load_mtp_moe_ffn(
    gguf: &GgufFile,
    cfg: &Qwen35Config,
    layer_index: u32,
    p: &str,
    device: &MlxDevice,
) -> Result<(MtpFfnWeightsGpu, u32)> {
    let moe_cfg = cfg.moe.as_ref().ok_or_else(|| {
        anyhow::anyhow!(
            "qwen35 MTP loader: block {layer_index} has MoE FFN tensors (`{p}.ffn_gate_exps.weight`) \
             but `cfg.moe` is None — the model metadata is inconsistent"
        )
    })?;
    // Load with the same quantized path used by every other MoE layer; this
    // keeps expert weights as native GGML blocks (no F32 expansion) and
    // matches what the verifier's main forward path consumes.
    let weights_q = load_moe_ffn_quantized(gguf, layer_index, device)
        .with_context(|| format!("MTP MoE FFN layer {layer_index}"))?;
    let moe_gpu = MoeFfnWeightsGpuQ::from_quantized(
        weights_q.expert_gate_q.clone(),
        weights_q.expert_up_q.clone(),
        weights_q.expert_down_q.clone(),
        weights_q.ggml_type_gate_up,
        weights_q.ggml_type_down,
        moe_cfg.num_experts,
        moe_cfg.moe_intermediate_size,
        cfg.hidden_size,
        &weights_q.router,
        &weights_q.shared_gate_logit,
        &weights_q.shared_gate,
        &weights_q.shared_up,
        &weights_q.shared_down,
        device,
    )
    .with_context(|| format!("MTP MoE upload layer {layer_index}"))?;
    let shape = MoeFfnShape {
        hidden_size: cfg.hidden_size,
        num_experts: moe_cfg.num_experts,
        num_experts_per_tok: moe_cfg.num_experts_per_tok,
        moe_intermediate_size: moe_cfg.moe_intermediate_size,
        shared_intermediate_size: moe_cfg.shared_expert_intermediate_size,
    };
    Ok((
        MtpFfnWeightsGpu::Moe {
            weights: moe_gpu,
            shape,
        },
        moe_cfg.moe_intermediate_size,
    ))
}

fn load_norm_gpu(gguf: &GgufFile, name: &str, len: usize, device: &MlxDevice) -> Result<MlxBuffer> {
    let data = load_f32_tensor(gguf, name, device).with_context(|| name.to_string())?;
    ensure!(data.len() == len, "{name} length {} != {len}", data.len());
    // W-5b.7 iter 2: MTP norm weights are static / reused across decode tokens —
    // route them through the residency-aware helper.
    upload_f32_weight(&data, device).with_context(|| format!("upload {name}"))
}