// arch/whisper/weights.cpp - read_whisper_hparams + build_whisper_weights.
// Read every required KV explicitly, validate cross-field invariants, then
// resolve each tensor slot against the GGUF with expected shape.
//
// Whisper notes: q/v/out carry bias, k does NOT (no "attn.k.bias" slot, both
// self and cross); logits head has no separate lm_head, the decoder reuses
// dec.token_embd.weight (tied); suppress_tokens / begin_suppress_tokens are
// int32 arrays (empty for .en variants); normalize tag "whisper_logmel" =
// per-utterance log10(max(mel, 1e-10)) -> clamp(max-8) -> (+4)/4.

#include "weights.h"

#include "ggml.h"
#include "gguf.h"
#include "transcribe-log.h"
#include "transcribe-meta.h"
#include "transcribe-weights-util.h"

#include <cstdio>
#include <cstring>
#include <initializer_list>

namespace transcribe::whisper {

namespace {

constexpr const char * kFamilyTag = "whisper";

// Optional u32 KV. Absent leaves `out` untouched; BadType is fatal.
transcribe_status read_optional_u32_kv(const gguf_context * gguf,
                                       const char *         key,
                                       const char *         error_tag,
                                       int32_t &            out) {
    uint32_t   value = 0;
    const auto st    = transcribe::read_uint32_kv(gguf, key, value);
    if (st == transcribe::KvResult::Ok) {
        out = static_cast<int32_t>(value);
        return TRANSCRIBE_OK;
    }
    if (st == transcribe::KvResult::Absent) {
        return TRANSCRIBE_OK;
    }
    log_msg(TRANSCRIBE_LOG_LEVEL_ERROR, "%s: KV %s has wrong type", error_tag, key);
    return TRANSCRIBE_ERR_GGUF;
}

// Optional int32 array KV. Absent leaves `out` untouched.
transcribe_status read_optional_i32_array_kv(const gguf_context *   gguf,
                                             const char *           key,
                                             const char *           error_tag,
                                             std::vector<int32_t> & out) {
    std::vector<int32_t> tmp;
    const auto           st = transcribe::read_int32_array_kv(gguf, key, tmp);
    if (st == transcribe::KvResult::Ok) {
        out = std::move(tmp);
        return TRANSCRIBE_OK;
    }
    if (st == transcribe::KvResult::Absent) {
        return TRANSCRIBE_OK;
    }
    log_msg(TRANSCRIBE_LOG_LEVEL_ERROR, "%s: KV %s has wrong type (expected int32 array)", error_tag, key);
    return TRANSCRIBE_ERR_GGUF;
}

// Optional float32 KV.
transcribe_status read_optional_f32_kv(const gguf_context * gguf,
                                       const char *         key,
                                       const char *         error_tag,
                                       float                default_value,
                                       float &              out) {
    float      value = default_value;
    const auto st    = transcribe::read_float32_kv(gguf, key, value);
    if (st == transcribe::KvResult::Ok) {
        out = value;
        return TRANSCRIBE_OK;
    }
    if (st == transcribe::KvResult::Absent) {
        out = default_value;
        return TRANSCRIBE_OK;
    }
    log_msg(TRANSCRIBE_LOG_LEVEL_ERROR, "%s: KV %s has wrong type", error_tag, key);
    return TRANSCRIBE_ERR_GGUF;
}

}  // namespace

transcribe_status read_whisper_hparams(const gguf_context * gguf, WhisperHParams & hp) {
    if (gguf == nullptr) {
        return TRANSCRIBE_ERR_INVALID_ARG;
    }

    // Encoder.
    if (auto st = read_required_u32_kv(gguf, "stt.whisper.encoder.n_layers", kFamilyTag, hp.enc_n_layers);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_required_u32_kv(gguf, "stt.whisper.encoder.d_model", kFamilyTag, hp.enc_d_model);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_required_u32_kv(gguf, "stt.whisper.encoder.n_heads", kFamilyTag, hp.enc_n_heads);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_required_u32_kv(gguf, "stt.whisper.encoder.ffn_dim", kFamilyTag, hp.enc_ffn_dim);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_required_u32_kv(gguf, "stt.whisper.encoder.num_mel_bins", kFamilyTag, hp.enc_num_mel_bins);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_required_u32_kv(gguf, "stt.whisper.encoder.max_source_positions", kFamilyTag,
                                       hp.enc_max_source_positions);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_required_string_kv(gguf, "stt.whisper.encoder.activation", kFamilyTag, hp.enc_activation);
        st != TRANSCRIBE_OK) {
        return st;
    }

    // Decoder.
    if (auto st = read_required_u32_kv(gguf, "stt.whisper.decoder.n_layers", kFamilyTag, hp.dec_n_layers);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_required_u32_kv(gguf, "stt.whisper.decoder.d_model", kFamilyTag, hp.dec_d_model);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_required_u32_kv(gguf, "stt.whisper.decoder.n_heads", kFamilyTag, hp.dec_n_heads);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_required_u32_kv(gguf, "stt.whisper.decoder.ffn_dim", kFamilyTag, hp.dec_ffn_dim);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_required_u32_kv(gguf, "stt.whisper.decoder.max_target_positions", kFamilyTag,
                                       hp.dec_max_target_positions);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_required_u32_kv(gguf, "stt.whisper.decoder.vocab_size", kFamilyTag, hp.dec_vocab_size);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_required_string_kv(gguf, "stt.whisper.decoder.activation", kFamilyTag, hp.dec_activation);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_optional_bool_kv(gguf, "stt.whisper.decoder.tie_word_embeddings", kFamilyTag, true,
                                        hp.dec_tie_word_embeddings);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_optional_bool_kv(gguf, "stt.whisper.decoder.scale_embedding", kFamilyTag, false,
                                        hp.dec_scale_embedding);
        st != TRANSCRIBE_OK) {
        return st;
    }

    // Whisper generation contract.
    if (auto st =
            read_required_u32_kv(gguf, "stt.whisper.decoder_start_token_id", kFamilyTag, hp.decoder_start_token_id);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st =
            read_required_u32_kv(gguf, "stt.whisper.no_timestamps_token_id", kFamilyTag, hp.no_timestamps_token_id);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_optional_u32_kv(gguf, "stt.whisper.sot_token_id", kFamilyTag, hp.sot_token_id);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_optional_u32_kv(gguf, "stt.whisper.transcribe_token_id", kFamilyTag, hp.transcribe_token_id);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_optional_u32_kv(gguf, "stt.whisper.translate_token_id", kFamilyTag, hp.translate_token_id);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_optional_u32_kv(gguf, "stt.whisper.prev_sot_token_id", kFamilyTag, hp.prev_sot_token_id);
        st != TRANSCRIBE_OK) {
        return st;
    }

    // sot falls back to decoder_start when unspecified — whisper-multilingual
    // uses the same id for both.
    if (hp.sot_token_id < 0) {
        hp.sot_token_id = hp.decoder_start_token_id;
    }

    if (auto st = read_optional_i32_array_kv(gguf, "stt.whisper.suppress_tokens", kFamilyTag, hp.suppress_tokens);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st =
            read_optional_i32_array_kv(gguf, "stt.whisper.begin_suppress_tokens", kFamilyTag, hp.begin_suppress_tokens);
        st != TRANSCRIBE_OK) {
        return st;
    }

    // Frontend.
    if (auto st = read_required_string_kv(gguf, "stt.frontend.type", kFamilyTag, hp.fe_type); st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_required_u32_kv(gguf, "stt.frontend.num_mels", kFamilyTag, hp.fe_num_mels);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_required_u32_kv(gguf, "stt.frontend.sample_rate", kFamilyTag, hp.fe_sample_rate);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_required_u32_kv(gguf, "stt.frontend.n_fft", kFamilyTag, hp.fe_n_fft); st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_required_u32_kv(gguf, "stt.frontend.win_length", kFamilyTag, hp.fe_win_length);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_required_u32_kv(gguf, "stt.frontend.hop_length", kFamilyTag, hp.fe_hop_length);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_required_string_kv(gguf, "stt.frontend.window", kFamilyTag, hp.fe_window); st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_required_string_kv(gguf, "stt.frontend.normalize", kFamilyTag, hp.fe_normalize);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_optional_f32_kv(gguf, "stt.frontend.dither", kFamilyTag, 0.0f, hp.fe_dither);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_optional_f32_kv(gguf, "stt.frontend.pre_emphasis", kFamilyTag, 0.0f, hp.fe_pre_emphasis);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_optional_f32_kv(gguf, "stt.frontend.f_min", kFamilyTag, 0.0f, hp.fe_f_min);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_optional_f32_kv(gguf, "stt.frontend.f_max", kFamilyTag, 8000.0f, hp.fe_f_max);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_optional_string_kv(gguf, "stt.frontend.pad_mode", kFamilyTag, "reflect", hp.fe_pad_mode);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_optional_bool_kv(gguf, "stt.frontend.center", kFamilyTag, true, hp.fe_center);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_optional_string_kv(gguf, "stt.frontend.mel_norm", kFamilyTag, "slaney", hp.fe_mel_norm);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_optional_u32_kv(gguf, "stt.frontend.chunk_length", kFamilyTag, hp.fe_chunk_length);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_optional_u32_kv(gguf, "stt.frontend.n_samples", kFamilyTag, hp.fe_n_samples);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_optional_u32_kv(gguf, "stt.frontend.nb_max_frames", kFamilyTag, hp.fe_nb_max_frames);
        st != TRANSCRIBE_OK) {
        return st;
    }

    // Capability flags. stt.capability.* is ecosystem-wide; the shared
    // read_capability_kv() (called later in the load path) re-reads
    // these into the public transcribe_capabilities struct, but we also
    // cache them here for easy access in decoder logic (supports_translate
    // gates the <|translate|> token availability etc).
    if (auto st = read_optional_bool_kv(gguf, "stt.capability.lang_detect", kFamilyTag, false, hp.cap_lang_detect);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_optional_bool_kv(gguf, "stt.capability.translate", kFamilyTag, false, hp.cap_translate);
        st != TRANSCRIBE_OK) {
        return st;
    }
    if (auto st = read_optional_bool_kv(gguf, "stt.capability.timestamps", kFamilyTag, false, hp.cap_timestamps);
        st != TRANSCRIBE_OK) {
        return st;
    }

    // Cross-field invariants.
    if (hp.enc_n_layers <= 0 || hp.enc_d_model <= 0 || hp.enc_n_heads <= 0 || hp.enc_ffn_dim <= 0 ||
        hp.enc_num_mel_bins <= 0 || hp.enc_max_source_positions <= 0) {
        log_msg(TRANSCRIBE_LOG_LEVEL_ERROR, "whisper: encoder hparams must be positive");
        return TRANSCRIBE_ERR_GGUF;
    }
    if (hp.enc_d_model % hp.enc_n_heads != 0) {
        log_msg(TRANSCRIBE_LOG_LEVEL_ERROR, "whisper: encoder d_model (%d) not divisible by n_heads (%d)",
                hp.enc_d_model, hp.enc_n_heads);
        return TRANSCRIBE_ERR_GGUF;
    }
    if (hp.dec_n_layers <= 0 || hp.dec_d_model <= 0 || hp.dec_n_heads <= 0 || hp.dec_ffn_dim <= 0 ||
        hp.dec_max_target_positions <= 0 || hp.dec_vocab_size <= 0) {
        log_msg(TRANSCRIBE_LOG_LEVEL_ERROR, "whisper: decoder hparams must be positive");
        return TRANSCRIBE_ERR_GGUF;
    }
    if (hp.dec_d_model % hp.dec_n_heads != 0) {
        log_msg(TRANSCRIBE_LOG_LEVEL_ERROR, "whisper: decoder d_model (%d) not divisible by n_heads (%d)",
                hp.dec_d_model, hp.dec_n_heads);
        return TRANSCRIBE_ERR_GGUF;
    }
    if (hp.enc_d_model != hp.dec_d_model) {
        // Upstream whisper always uses d_model=d_model; a mismatch would
        // break the cross-attention K/V dim contract.
        log_msg(TRANSCRIBE_LOG_LEVEL_ERROR,
                "whisper: encoder d_model (%d) != decoder d_model (%d); "
                "cross-attention would mismatch",
                hp.enc_d_model, hp.dec_d_model);
        return TRANSCRIBE_ERR_GGUF;
    }
    if (hp.enc_activation != "gelu" || hp.dec_activation != "gelu") {
        log_msg(TRANSCRIBE_LOG_LEVEL_ERROR,
                "whisper: only \"gelu\" activation is supported "
                "(enc=\"%s\", dec=\"%s\")",
                hp.enc_activation.c_str(), hp.dec_activation.c_str());
        return TRANSCRIBE_ERR_GGUF;
    }
    if (!hp.dec_tie_word_embeddings) {
        // Whisper always ties lm_head to token_embedding; a false here
        // would mean the converter shipped a separate head tensor we
        // don't currently load.
        log_msg(TRANSCRIBE_LOG_LEVEL_ERROR,
                "whisper: stt.whisper.decoder.tie_word_embeddings=false "
                "is not supported (no separate lm_head tensor in the "
                "catalog)");
        return TRANSCRIBE_ERR_GGUF;
    }
    if (hp.dec_scale_embedding) {
        // HF WhisperDecoder multiplies inputs_embeds by sqrt(d_model) when
        // config.scale_embedding=True. Upstream Whisper always ships False;
        // if this ever flips we need to add the scale to the decoder graph
        // (both prefill and KV paths, after get_rows(token_embd, ...)).
        log_msg(TRANSCRIBE_LOG_LEVEL_ERROR,
                "whisper: stt.whisper.decoder.scale_embedding=true is "
                "not supported (decoder graph does not apply the "
                "sqrt(d_model) embed scale)");
        return TRANSCRIBE_ERR_GGUF;
    }
    if (hp.decoder_start_token_id < 0 || hp.no_timestamps_token_id < 0) {
        log_msg(TRANSCRIBE_LOG_LEVEL_ERROR,
                "whisper: decoder_start_token_id / no_timestamps_token_id "
                "must be set");
        return TRANSCRIBE_ERR_GGUF;
    }
    if (hp.fe_num_mels <= 0 || hp.fe_sample_rate <= 0 || hp.fe_n_fft <= 0 || hp.fe_win_length <= 0 ||
        hp.fe_hop_length <= 0) {
        log_msg(TRANSCRIBE_LOG_LEVEL_ERROR, "whisper: frontend dimensions must be positive");
        return TRANSCRIBE_ERR_GGUF;
    }
    if (hp.fe_win_length != hp.fe_n_fft) {
        // Whisper uses a full-length periodic Hann window (win=n_fft=400).
        // A mismatch would require a windowed-shorter-than-FFT code path
        // that we don't implement.
        log_msg(TRANSCRIBE_LOG_LEVEL_ERROR,
                "whisper: frontend win_length (%d) != n_fft (%d); "
                "only full-length window is supported",
                hp.fe_win_length, hp.fe_n_fft);
        return TRANSCRIBE_ERR_GGUF;
    }
    if (hp.fe_num_mels != hp.enc_num_mel_bins) {
        log_msg(TRANSCRIBE_LOG_LEVEL_ERROR, "whisper: frontend num_mels (%d) != encoder num_mel_bins (%d)",
                hp.fe_num_mels, hp.enc_num_mel_bins);
        return TRANSCRIBE_ERR_GGUF;
    }
    if (hp.fe_type != "mel") {
        log_msg(TRANSCRIBE_LOG_LEVEL_ERROR, "whisper: unsupported frontend type \"%s\" (only \"mel\")",
                hp.fe_type.c_str());
        return TRANSCRIBE_ERR_GGUF;
    }
    if (hp.fe_window != "hann" && hp.fe_window != "hann_periodic") {
        log_msg(TRANSCRIBE_LOG_LEVEL_ERROR,
                "whisper: unsupported frontend window \"%s\" "
                "(only \"hann\"/\"hann_periodic\" is supported)",
                hp.fe_window.c_str());
        return TRANSCRIBE_ERR_GGUF;
    }
    if (hp.fe_normalize != "whisper_logmel") {
        log_msg(TRANSCRIBE_LOG_LEVEL_ERROR,
                "whisper: unsupported frontend normalize \"%s\" "
                "(only \"whisper_logmel\" is supported)",
                hp.fe_normalize.c_str());
        return TRANSCRIBE_ERR_GGUF;
    }
    if (hp.fe_mel_norm != "slaney") {
        log_msg(TRANSCRIBE_LOG_LEVEL_ERROR,
                "whisper: unsupported frontend mel_norm \"%s\" "
                "(only \"slaney\" is supported)",
                hp.fe_mel_norm.c_str());
        return TRANSCRIBE_ERR_GGUF;
    }
    if (hp.fe_pad_mode != "reflect") {
        log_msg(TRANSCRIBE_LOG_LEVEL_ERROR,
                "whisper: unsupported frontend pad_mode \"%s\" "
                "(only \"reflect\" is supported)",
                hp.fe_pad_mode.c_str());
        return TRANSCRIBE_ERR_GGUF;
    }
    if (!hp.fe_center) {
        log_msg(TRANSCRIBE_LOG_LEVEL_ERROR, "whisper: non-centered STFT is not supported");
        return TRANSCRIBE_ERR_GGUF;
    }
    if (hp.fe_dither != 0.0f) {
        // HF WhisperFeatureExtractor defaults dither=0.0 and the C++ mel
        // frontend never applies dither (transcribe-mel.cpp's contract).
        // A non-zero value in the GGUF would be silently dropped, so
        // reject it at load time rather than producing a silently-wrong
        // spectrogram.
        log_msg(TRANSCRIBE_LOG_LEVEL_ERROR,
                "whisper: stt.frontend.dither=%g is not supported "
                "(frontend is deterministic; expected 0.0)",
                hp.fe_dither);
        return TRANSCRIBE_ERR_GGUF;
    }

    return TRANSCRIBE_OK;
}

// Frontend (mel filterbank + Hann window install)
transcribe_status install_mel_from_buffers(const WhisperHParams &                   hp,
                                           std::vector<float>                       filterbank,
                                           std::vector<float>                       window,
                                           std::optional<transcribe::MelFrontend> & out_mel) {
    transcribe::MelConfig cfg{};
    cfg.sample_rate  = hp.fe_sample_rate;
    cfg.num_mels     = hp.fe_num_mels;
    cfg.n_fft        = hp.fe_n_fft;
    cfg.win_length   = hp.fe_win_length;
    cfg.hop_length   = hp.fe_hop_length;
    cfg.pre_emphasis = hp.fe_pre_emphasis;
    cfg.f_min        = hp.fe_f_min;
    cfg.f_max        = hp.fe_f_max;
    cfg.pad_mode     = hp.fe_pad_mode;  // "reflect"
    cfg.window_type  = hp.fe_window;    // "hann_periodic"

    // Map whisper's "whisper_logmel" tag to MelFrontend's
    // "per_utterance" mode, which implements exactly the
    // log10 → clamp(max-8) → (+4)/4 sequence. Other normalize
    // strings would be a converter bug — guard at load.
    if (hp.fe_normalize == "whisper_logmel" || hp.fe_normalize == "per_utterance") {
        cfg.normalize = "per_utterance";
    } else {
        log_msg(TRANSCRIBE_LOG_LEVEL_ERROR, "whisper: unsupported fe_normalize='%s'", hp.fe_normalize.c_str());
        return TRANSCRIBE_ERR_GGUF;
    }

    cfg.filterbank = std::move(filterbank);
    cfg.window     = std::move(window);

    out_mel.emplace(cfg);
    return TRANSCRIBE_OK;
}

// Weights
namespace {

using transcribe::weights::lname;

constexpr const char * kTag = kFamilyTag;

#define GET_F32(slot, name, ...)                                                                          \
    do {                                                                                                  \
        ggml_tensor * _t =                                                                                \
            transcribe::weights::find_tensor(ctx_meta, (name), { GGML_TYPE_F32 }, { __VA_ARGS__ }, kTag); \
        if (_t == nullptr)                                                                                \
            return TRANSCRIBE_ERR_GGUF;                                                                   \
        (slot) = _t;                                                                                      \
    } while (0)

#define GET_CONV(slot, name, ...)                                                                              \
    do {                                                                                                       \
        ggml_tensor * _t = transcribe::weights::find_tensor(ctx_meta, (name), { TRANSCRIBE_QUANT_CONV_TYPES }, \
                                                            { __VA_ARGS__ }, kTag);                            \
        if (_t == nullptr)                                                                                     \
            return TRANSCRIBE_ERR_GGUF;                                                                        \
        (slot) = _t;                                                                                           \
    } while (0)

#define GET_LIN(slot, name, ...)                                                                                 \
    do {                                                                                                         \
        ggml_tensor * _t = transcribe::weights::find_tensor(ctx_meta, (name), { TRANSCRIBE_QUANT_LINEAR_TYPES }, \
                                                            { __VA_ARGS__ }, kTag);                              \
        if (_t == nullptr)                                                                                       \
            return TRANSCRIBE_ERR_GGUF;                                                                          \
        (slot) = _t;                                                                                             \
    } while (0)

}  // namespace

transcribe_status build_whisper_weights(ggml_context * ctx_meta, const WhisperHParams & hp, WhisperWeights & weights) {
    if (ctx_meta == nullptr) {
        return TRANSCRIBE_ERR_INVALID_ARG;
    }

    const int64_t d_model    = hp.enc_d_model;
    const int64_t n_mel      = hp.enc_num_mel_bins;
    const int64_t enc_ff     = hp.enc_ffn_dim;
    const int64_t src_pos    = hp.enc_max_source_positions;
    const int64_t dec_h      = hp.dec_d_model;
    const int64_t dec_ff     = hp.dec_ffn_dim;
    const int64_t tgt_pos    = hp.dec_max_target_positions;
    const int64_t vocab_size = hp.dec_vocab_size;
    const int64_t n_fft_half = hp.fe_n_fft / 2 + 1;

    // ----- frontend (mel filterbank + hann window) -----
    GET_F32(weights.frontend.mel_filterbank, "frontend.mel_filterbank", n_fft_half, n_mel);
    GET_F32(weights.frontend.window, "frontend.window", hp.fe_n_fft);

    // ----- encoder conv stem -----
    // Conv1d kernels: PyTorch [out, in, K] -> ggml ne=[K, in, out].
    GET_CONV(weights.enc_stem.conv0_w, "enc.conv.0.weight", 3, n_mel, d_model);
    GET_F32(weights.enc_stem.conv0_b, "enc.conv.0.bias", d_model);
    GET_CONV(weights.enc_stem.conv1_w, "enc.conv.1.weight", 3, d_model, d_model);
    GET_F32(weights.enc_stem.conv1_b, "enc.conv.1.bias", d_model);

    // ----- encoder top (pos emb + final LN) -----
    // pos_emb is nn.Embedding weight [max_source_positions, d_model] ->
    // ggml ne = [d_model, max_source_positions].
    GET_F32(weights.enc_top.pos_emb_w, "enc.pos_emb.weight", d_model, src_pos);
    GET_F32(weights.enc_top.final_norm_w, "enc.final_norm.weight", d_model);
    GET_F32(weights.enc_top.final_norm_b, "enc.final_norm.bias", d_model);

    // ----- encoder blocks -----
    weights.enc_blocks.assign(hp.enc_n_layers, WhisperEncBlock{});
    for (int i = 0; i < hp.enc_n_layers; ++i) {
        auto & b = weights.enc_blocks[i];

        GET_F32(b.norm_attn_w, lname("enc.blocks.%d.norm_attn.weight", i), d_model);
        GET_F32(b.norm_attn_b, lname("enc.blocks.%d.norm_attn.bias", i), d_model);

        GET_LIN(b.attn_q_w, lname("enc.blocks.%d.attn.q.weight", i), d_model, d_model);
        GET_F32(b.attn_q_b, lname("enc.blocks.%d.attn.q.bias", i), d_model);
        GET_LIN(b.attn_k_w, lname("enc.blocks.%d.attn.k.weight", i), d_model, d_model);  // no bias
        GET_LIN(b.attn_v_w, lname("enc.blocks.%d.attn.v.weight", i), d_model, d_model);
        GET_F32(b.attn_v_b, lname("enc.blocks.%d.attn.v.bias", i), d_model);
        GET_LIN(b.attn_out_w, lname("enc.blocks.%d.attn.out.weight", i), d_model, d_model);
        GET_F32(b.attn_out_b, lname("enc.blocks.%d.attn.out.bias", i), d_model);

        GET_F32(b.norm_ffn_w, lname("enc.blocks.%d.norm_ffn.weight", i), d_model);
        GET_F32(b.norm_ffn_b, lname("enc.blocks.%d.norm_ffn.bias", i), d_model);
        GET_LIN(b.ffn_fc1_w, lname("enc.blocks.%d.ffn.fc1.weight", i), d_model, enc_ff);
        GET_F32(b.ffn_fc1_b, lname("enc.blocks.%d.ffn.fc1.bias", i), enc_ff);
        GET_LIN(b.ffn_fc2_w, lname("enc.blocks.%d.ffn.fc2.weight", i), enc_ff, d_model);
        GET_F32(b.ffn_fc2_b, lname("enc.blocks.%d.ffn.fc2.bias", i), d_model);
    }

    // ----- decoder top -----
    // Token embedding: PyTorch [vocab, dec_h] -> ggml ne = [dec_h, vocab].
    // Doubles as the tied lm_head weight, so it accepts the full quant
    // allowlist (GET_LIN, not GET_F32).
    GET_LIN(weights.dec_top.token_embd_w, "dec.token_embd.weight", dec_h, vocab_size);
    GET_F32(weights.dec_top.pos_emb_w, "dec.pos_emb.weight", dec_h, tgt_pos);
    GET_F32(weights.dec_top.final_norm_w, "dec.final_norm.weight", dec_h);
    GET_F32(weights.dec_top.final_norm_b, "dec.final_norm.bias", dec_h);

    // ----- decoder blocks -----
    weights.dec_blocks.assign(hp.dec_n_layers, WhisperDecBlock{});
    for (int i = 0; i < hp.dec_n_layers; ++i) {
        auto & b = weights.dec_blocks[i];

        // Self-attention.
        GET_F32(b.norm_self_w, lname("dec.blocks.%d.norm_self.weight", i), dec_h);
        GET_F32(b.norm_self_b, lname("dec.blocks.%d.norm_self.bias", i), dec_h);
        GET_LIN(b.self_q_w, lname("dec.blocks.%d.self_attn.q.weight", i), dec_h, dec_h);
        GET_F32(b.self_q_b, lname("dec.blocks.%d.self_attn.q.bias", i), dec_h);
        GET_LIN(b.self_k_w, lname("dec.blocks.%d.self_attn.k.weight", i), dec_h, dec_h);  // no bias
        GET_LIN(b.self_v_w, lname("dec.blocks.%d.self_attn.v.weight", i), dec_h, dec_h);
        GET_F32(b.self_v_b, lname("dec.blocks.%d.self_attn.v.bias", i), dec_h);
        GET_LIN(b.self_out_w, lname("dec.blocks.%d.self_attn.out.weight", i), dec_h, dec_h);
        GET_F32(b.self_out_b, lname("dec.blocks.%d.self_attn.out.bias", i), dec_h);

        // Cross-attention.
        GET_F32(b.norm_cross_w, lname("dec.blocks.%d.norm_cross.weight", i), dec_h);
        GET_F32(b.norm_cross_b, lname("dec.blocks.%d.norm_cross.bias", i), dec_h);
        GET_LIN(b.cross_q_w, lname("dec.blocks.%d.cross_attn.q.weight", i), dec_h, dec_h);
        GET_F32(b.cross_q_b, lname("dec.blocks.%d.cross_attn.q.bias", i), dec_h);
        GET_LIN(b.cross_k_w, lname("dec.blocks.%d.cross_attn.k.weight", i), dec_h, dec_h);  // no bias
        GET_LIN(b.cross_v_w, lname("dec.blocks.%d.cross_attn.v.weight", i), dec_h, dec_h);
        GET_F32(b.cross_v_b, lname("dec.blocks.%d.cross_attn.v.bias", i), dec_h);
        GET_LIN(b.cross_out_w, lname("dec.blocks.%d.cross_attn.out.weight", i), dec_h, dec_h);
        GET_F32(b.cross_out_b, lname("dec.blocks.%d.cross_attn.out.bias", i), dec_h);

        // FFN.
        GET_F32(b.norm_ffn_w, lname("dec.blocks.%d.norm_ffn.weight", i), dec_h);
        GET_F32(b.norm_ffn_b, lname("dec.blocks.%d.norm_ffn.bias", i), dec_h);
        GET_LIN(b.ffn_fc1_w, lname("dec.blocks.%d.ffn.fc1.weight", i), dec_h, dec_ff);
        GET_F32(b.ffn_fc1_b, lname("dec.blocks.%d.ffn.fc1.bias", i), dec_ff);
        GET_LIN(b.ffn_fc2_w, lname("dec.blocks.%d.ffn.fc2.weight", i), dec_ff, dec_h);
        GET_F32(b.ffn_fc2_b, lname("dec.blocks.%d.ffn.fc2.bias", i), dec_h);
    }

    return TRANSCRIBE_OK;
}

#undef GET_F32
#undef GET_CONV
#undef GET_LIN

}  // namespace transcribe::whisper