Skip to content

LLM inference in C/C++

C++ ggml

model : add Tencent Hy 4 (hy_v4) preview architecture support (#28127)

HongHuang49c0dc82b849
diff --git a/conversion/__init__.py b/conversion/__init__.py
index ba73192efa..94d6a49fbc 100644
--- a/conversion/__init__.py
+++ b/conversion/__init__.py
@@ -124,6 +124,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
"HunYuanMoEV1ForCausalLM": "hunyuan",
"HunYuanVLForConditionalGeneration": "hunyuan",
"HYV3ForCausalLM": "hunyuan",
+ "HYV4ForCausalLM": "hy_v4",
"IQuestCoderForCausalLM": "llama",
"InternLM2ForCausalLM": "internlm",
"InternLM3ForCausalLM": "internlm",
diff --git a/conversion/base.py b/conversion/base.py
index daae28e92a..c1ecf1c651 100644
--- a/conversion/base.py
+++ b/conversion/base.py
@@ -1507,6 +1507,9 @@ class TextModel(ModelBase):
if chkhsh == "bba3b3366b646dbdded5dbc42d59598b849371afc42f7beafa914afaa5b70aa6":
# ref: https://huggingface.co/tencent/Hunyuan-4B-Instruct
res = "hunyuan-dense"
+ if chkhsh == "e6ddf9c6686791c12d698d34c31ab9be1fea9af5a3d9a6909783ab382198ae1c":
+ # ref: https://huggingface.co/tencent/Hy4-preview
+ res = "hy_v4"
if chkhsh == "a6b57017d60e6edb4d88ecc2845188e0eb333a70357e45dcc9b53964a73bbae6":
# ref: https://huggingface.co/tiiuae/Falcon-H1-0.5B-Base
res = "falcon-h1"
diff --git a/conversion/hy_v4.py b/conversion/hy_v4.py
new file mode 100644
index 0000000000..f564b9ec25
--- /dev/null
+++ b/conversion/hy_v4.py
@@ -0,0 +1,311 @@
+from __future__ import annotations
+
+import re
+from typing import Iterable
+
+import torch
+
+from .base import ModelBase, gguf, logger
+from .deepseek import DeepseekV2Model
+
+
+def split_kv_b_proj(weight: torch.Tensor, n_head: int, qk_nope: int, v_head_dim: int):
+ """Split kv_b_proj into k_b (transposed) and v_b, matching DeepSeek MLA absorption.
+
+ weight: [n_head*(qk_nope+v_head_dim), kv_lora_rank].
+ Returns (k_b, v_b): k_b [n_head, kv_lora_rank, qk_nope], v_b [n_head, v_head_dim, kv_lora_rank].
+ """
+ kv_lora = weight.shape[-1]
+ assert weight.shape[0] == n_head * (qk_nope + v_head_dim)
+ kv_b = weight.view(n_head, qk_nope + v_head_dim, kv_lora)
+ k_b, v_b = torch.split(kv_b, [qk_nope, v_head_dim], dim=1)
+ k_b = k_b.transpose(1, 2).contiguous() # [n_head, kv_lora, qk_nope]
+ return k_b, v_b.contiguous()
+
+
+def split_gate_up(weight: torch.Tensor, moe_intermediate_size: int):
+ """Split a fused stacked gate_up expert tensor into (gate, up).
+
+ weight: [n_expert, 2*moe_intermediate_size, hidden] (gate first, up second).
+ Returns (gate, up) each [n_expert, moe_intermediate_size, hidden].
+ """
+ assert weight.shape[1] == 2 * moe_intermediate_size, f"{weight.shape[1]} != 2*{moe_intermediate_size}"
+ gate = weight[:, :moe_intermediate_size, :].contiguous()
+ up = weight[:, moe_intermediate_size:, :].contiguous()
+ return gate, up
+
+
[email protected]("HYV4ForCausalLM")
+class HYV4Model(DeepseekV2Model):
+ """HY_V4: DeepSeek-V3 style MLA + MoE with iHC, a gated MLA output and a learnable sink.
+
+ Reuses DeepseekV2Model for the vocab and the MLA metadata, but overrides the tensor mapping
+ because HY_V4 ships pre-stacked / fused experts plus extra iHC, gate and sink tensors. The
+ rope rows are mapped straight through (no permute) - the graph rotates consecutive pairs.
+
+ DSA is supported: indexer weights are exported for the layers marked "full" in indexer_types.
+ "shared" layers reuse the top-k of the last preceding full layer at inference time, so they
+ carry no indexer weights.
+
+ MTP (num_nextn_predict_layers) is dropped, so the GGUF cannot be used for speculative
+ decoding. The reference only runs the MTP layers while training or while speculating, so they
+ cannot change single-token logits.
+ """
+
+ model_arch = gguf.MODEL_ARCH.HY_V4
+
+ # tensors a "full" indexer layer must carry
+ INDEXER_SUFFIXES = frozenset({
+ "self_attn.indexer.wq_b.weight",
+ "self_attn.indexer.wk.weight",
+ "self_attn.indexer.k_norm.weight",
+ "self_attn.indexer.k_norm.bias",
+ "self_attn.indexer.weights_proj.weight",
+ })
+
+ @classmethod
+ def filter_tensors(cls, item):
+ # drop MTP here, not in modify_tensors, so the weights are never read
+ if item[0].startswith("model.mtp_layers."):
+ return None
+ return super().filter_tensors(item)
+
+ def _check_indexer_hparams(self):
+ for key in ("index_n_heads", "index_head_dim", "index_topk"):
+ if key not in self.hparams:
+ raise ValueError(f"HY_V4 has DSA layers but no {key}")
+
+ def indexer_is_full(self) -> list[bool] | None:
+ """Per-layer indexer ownership, or None when the checkpoint has no DSA.
+
+ indexer_types entries are "full" (owns an indexer) or "shared" (reuses the preceding
+ full layer's top-k). Missing indexer_types with sparse layers means every sparse layer
+ owns one.
+ """
+ hparams = self.hparams
+ n_layer = hparams["num_hidden_layers"]
+ indexer_types = hparams.get("indexer_types")
+
+ # the reference drives DSA off indexer_types alone; layer_types is only a fallback for
+ # checkpoints predating it (it was renamed to deepseek_sparse_attention upstream)
+ if indexer_types is None:
+ layer_types = hparams.get("layer_types") or []
+ sparse = {"sparse_attention", "deepseek_sparse_attention"}
+ if not any(t in sparse for t in layer_types):
+ return None
+ if len(layer_types) < n_layer:
+ raise ValueError(f"HY_V4 layer_types has {len(layer_types)} entries, need {n_layer}")
+ self._check_indexer_hparams()
+ return [t in sparse for t in layer_types[:n_layer]]
+
+ self._check_indexer_hparams()
+
+ if len(indexer_types) < n_layer:
+ raise ValueError(f"HY_V4 indexer_types has {len(indexer_types)} entries, need {n_layer}")
+ unknown = {t for t in indexer_types[:n_layer]} - {"full", "shared"}
+ if unknown:
+ raise ValueError(f"HY_V4 unknown indexer_types values: {sorted(unknown)}")
+ is_full = [t == "full" for t in indexer_types[:n_layer]]
+ if is_full and not is_full[0]:
+ raise ValueError("HY_V4 layer 0 must be indexer_types 'full' (nothing precedes it to share)")
+ return is_full
+
+ def set_gguf_parameters(self):
+ hparams = self.hparams
+
+ # HY4 has n_group == topk_group == 1 (no group routing). Drop the keys so the base does
+ # not emit expert_group_count/used; llama.cpp then takes the ungrouped MoE path.
+ if hparams.get("n_group") == 1 and hparams.get("topk_group") == 1:
+ hparams.pop("n_group", None)
+ hparams.pop("topk_group", None)
+
+ # HY_V4 config expresses dense/sparse layers via mlp_layer_types, but DeepseekV2Model
+ # needs first_k_dense_replace. Derive it as the contiguous leading "dense" block
+ # (the real config.json also carries first_k_dense_replace; prefer it when present,
+ # but assert the two agree so a mismatch fails loudly).
+ mlp_types = hparams.get("mlp_layer_types")
+ explicit = hparams.get("first_k_dense_replace")
+ derived = None
+ if mlp_types is not None:
+ lead = 0
+ for t in mlp_types:
+ if t == "dense":
+ lead += 1
+ else:
+ break
+ if any(t == "dense" for t in mlp_types[lead:]):
+ raise NotImplementedError("HY_V4 converter expects a contiguous leading dense block")
+ derived = lead
+ if explicit is not None and derived is not None and explicit != derived:
+ raise ValueError(
+ f"HY_V4 first_k_dense_replace ({explicit}) disagrees with mlp_layer_types "
+ f"leading-dense count ({derived})"
+ )
+ if explicit is None:
+ if derived is None:
+ raise ValueError("HY_V4 needs first_k_dense_replace or mlp_layer_types to place dense layers")
+ hparams["first_k_dense_replace"] = derived
+
+ # reuse DeepseekV2 MLA + MoE metadata (forces num_key_value_heads=1, writes q/kv lora,
+ # key/value lengths, expert counts, weights scale/norm, rope dims, etc.)
+ super().set_gguf_parameters()
+
+ # HY4 uses DeepSeek-V3 sigmoid routing with e_score_correction_bias. The config has no
+ # scoring_func key, so the base does not write a gating func; set it explicitly.
+ self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
+
+ # routed-expert SwiGLU logits clamp (only routed experts; shared/dense are not clamped,
+ # so swiglu_clamp_shexp is intentionally not written). 0.0 disables the clamp.
+ swiglu_limit = float(hparams.get("swiglu_limit", 0.0) or 0.0)
+ if swiglu_limit > 0.0:
+ self.gguf_writer.add_swiglu_clamp_exp([swiglu_limit] * self.block_count)
+
+ # iHC (independent Hyper-Connections)
+ self.gguf_writer.add_hyper_connection_count(hparams["hc_mult"])
+ self.gguf_writer.add_hyper_connection_epsilon(hparams["hc_eps"])
+ self.gguf_writer.add_hyper_connection_magnitude(hparams["hc_magnitude"])
+
+ # is_full is written explicitly; the graph must not infer it from tensor presence
+ is_full = self.indexer_is_full()
+ if is_full is not None:
+ self.gguf_writer.add_indexer_head_count(hparams["index_n_heads"])
+ self.gguf_writer.add_indexer_key_length(hparams["index_head_dim"])
+ self.gguf_writer.add_indexer_top_k(hparams["index_topk"])
+ self.gguf_writer.add_indexer_types(is_full)
+ logger.info(
+ "HY_V4 DSA: %d/%d layers own an indexer (top_k=%d, n_heads=%d, head_dim=%d)",
+ sum(is_full), len(is_full), hparams["index_topk"],
+ hparams["index_n_heads"], hparams["index_head_dim"],
+ )
+
+ if hparams.get("num_nextn_predict_layers", 0):
+ logger.warning(
+ "HY_V4: dropping %d MTP (nextn) layer(s) - the reference runs them only under "
+ "training / speculative decoding. This GGUF cannot be used for speculative decoding.",
+ hparams["num_nextn_predict_layers"],
+ )
+
+ def prepare_tensors(self):
+ # validate before the base materializes tensors, so a mismatch fails early
+ is_full = self.indexer_is_full()
+ if is_full is not None:
+ present: dict[int, set[str]] = {}
+ for name in self.model_tensors:
+ m = re.match(r"model\.layers\.(\d+)\.(self_attn\.indexer\..+)$", name)
+ if m:
+ present.setdefault(int(m.group(1)), set()).add(m.group(2))
+ for il, expect_full in enumerate(is_full):
+ seen = present.get(il, set())
+ if expect_full and seen != self.INDEXER_SUFFIXES:
+ raise ValueError(
+ f"HY_V4 layer {il} is indexer_types 'full' but is missing indexer tensors: "
+ f"{sorted(self.INDEXER_SUFFIXES - seen)}"
+ )
+ if not expect_full and seen:
+ raise ValueError(
+ f"HY_V4 layer {il} is indexer_types 'shared' but carries indexer tensors: "
+ f"{sorted(seen)}"
+ )
+
+ super().prepare_tensors()
+
+ def tensor_force_quant(self, name, new_name, bid, n_dims):
+ # iHC mixing matrices are 2D .weight tensors that the reference keeps in fp32
+ # (_keep_in_fp32_modules_strict). 1D tensors (hc_base/scale, attn_sinks,
+ # e_score_correction_bias) and the router (FFN_GATE_INP) are already forced F32 by the
+ # base rules. Force the HC *_fn matrices here.
+ if new_name.endswith(("hc_attn_fn.weight", "hc_ffn_fn.weight", "output_hc_fn.weight")):
+ return gguf.GGMLQuantizationType.F32
+ # indexer k_norm is fp32 in the reference; the base rules already cover
+ # *_norm.weight and INDEXER_PROJ, but not this bias
+ if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.INDEXER_K_NORM, bid, suffix=".bias"):
+ return gguf.GGMLQuantizationType.F32
+ # enable_lm_head_fp32: mirror the reference fp32 LM-head matmul by keeping output F32.
+ if new_name == "output.weight" and self.hparams.get("enable_lm_head_fp32", False):
+ return gguf.GGMLQuantizationType.F32
+ return super().tensor_force_quant(name, new_name, bid, n_dims)
+
+ def modify_tensors(self, data_torch: torch.Tensor, name: str, bid: int | None) -> Iterable[tuple[str, torch.Tensor]]:
+ hparams = self.hparams
+ n_head = hparams["num_attention_heads"]
+ qk_nope = hparams["qk_nope_head_dim"]
+ v_head_dim = hparams["v_head_dim"]
+ moe_inter = hparams["moe_intermediate_size"]
+
+ tn = self.format_tensor_name
+
+ # ---- global (non per-layer) ----
+ if name == "model.embed_tokens.weight":
+ return [(tn(gguf.MODEL_TENSOR.TOKEN_EMBD), data_torch)]
+ if name == "model.norm.weight":
+ return [(tn(gguf.MODEL_TENSOR.OUTPUT_NORM), data_torch)]
+ if name == "lm_head.weight":
+ return [(tn(gguf.MODEL_TENSOR.OUTPUT), data_torch)]
+ if name == "model.hc_head.hc_head_fn":
+ return [(tn(gguf.MODEL_TENSOR.HC_HEAD_FN), data_torch)]
+ if name == "model.hc_head.hc_head_base":
+ return [(tn(gguf.MODEL_TENSOR.HC_HEAD_BASE), data_torch)]
+ if name == "model.hc_head.hc_head_scale":
+ return [(tn(gguf.MODEL_TENSOR.HC_HEAD_SCALE), data_torch)]
+
+ assert bid is not None, f"expected a per-layer tensor, got {name!r}"
+
+ # ---- per-layer, keyed by suffix after 'model.layers.{bid}.' ----
+ suffix = name.split(f"model.layers.{bid}.", 1)[-1]
+
+ # note: q_b_proj and kv_a_proj_with_mqa are mapped straight through (no RoPE permute),
+ # the graph rotates consecutive pairs so the rows need no reordering
+ simple = {
+ "input_layernorm.weight": (gguf.MODEL_TENSOR.ATTN_NORM, ".weight"),
+ "post_attention_layernorm.weight": (gguf.MODEL_TENSOR.FFN_NORM, ".weight"),
+ "self_attn.q_a_proj.weight": (gguf.MODEL_TENSOR.ATTN_Q_A, ".weight"),
+ "self_attn.q_a_layernorm.weight": (gguf.MODEL_TENSOR.ATTN_Q_A_NORM, ".weight"),
+ "self_attn.q_b_proj.weight": (gguf.MODEL_TENSOR.ATTN_Q_B, ".weight"),
+ "self_attn.kv_a_proj_with_mqa.weight": (gguf.MODEL_TENSOR.ATTN_KV_A_MQA, ".weight"),
+ "self_attn.kv_a_layernorm.weight": (gguf.MODEL_TENSOR.ATTN_KV_A_NORM, ".weight"),
+ "self_attn.o_proj.weight": (gguf.MODEL_TENSOR.ATTN_OUT, ".weight"),
+ "self_attn.linear_gate.weight": (gguf.MODEL_TENSOR.ATTN_GATE, ".weight"),
+ "self_attn.learnable_sink_param": (gguf.MODEL_TENSOR.ATTN_SINKS, ".weight"),
+ "self_attn.indexer.wq_b.weight": (gguf.MODEL_TENSOR.INDEXER_ATTN_Q_B, ".weight"),
+ "self_attn.indexer.wk.weight": (gguf.MODEL_TENSOR.INDEXER_ATTN_K, ".weight"),
+ "self_attn.indexer.k_norm.weight": (gguf.MODEL_TENSOR.INDEXER_K_NORM, ".weight"),
+ "self_attn.indexer.k_norm.bias": (gguf.MODEL_TENSOR.INDEXER_K_NORM, ".bias"),
+ "self_attn.indexer.weights_proj.weight": (gguf.MODEL_TENSOR.INDEXER_PROJ, ".weight"),
+ "hc_attn_layer.hc_pre.hc_fn": (gguf.MODEL_TENSOR.HC_ATTN_FN, ".weight"),
+ "hc_attn_layer.hc_pre.hc_base": (gguf.MODEL_TENSOR.HC_ATTN_BASE, ".weight"),
+ "hc_attn_layer.hc_pre.hc_scale": (gguf.MODEL_TENSOR.HC_ATTN_SCALE, ".weight"),
+ "hc_mlp_layer.hc_pre.hc_fn": (gguf.MODEL_TENSOR.HC_FFN_FN, ".weight"),
+ "hc_mlp_layer.hc_pre.hc_base": (gguf.MODEL_TENSOR.HC_FFN_BASE, ".weight"),
+ "hc_mlp_layer.hc_pre.hc_scale": (gguf.MODEL_TENSOR.HC_FFN_SCALE, ".weight"),
+ "mlp.gate.weight": (gguf.MODEL_TENSOR.FFN_GATE_INP, ".weight"),
+ "mlp.gate.e_score_correction.bias":(gguf.MODEL_TENSOR.FFN_EXP_PROBS_B, ".bias"),
+ "mlp.gate_proj.weight": (gguf.MODEL_TENSOR.FFN_GATE, ".weight"),
+ "mlp.up_proj.weight": (gguf.MODEL_TENSOR.FFN_UP, ".weight"),
+ "mlp.down_proj.weight": (gguf.MODEL_TENSOR.FFN_DOWN, ".weight"),
+ "mlp.shared_experts.gate_proj.weight": (gguf.MODEL_TENSOR.FFN_GATE_SHEXP, ".weight"),
+ "mlp.shared_experts.up_proj.weight": (gguf.MODEL_TENSOR.FFN_UP_SHEXP, ".weight"),
+ "mlp.shared_experts.down_proj.weight": (gguf.MODEL_TENSOR.FFN_DOWN_SHEXP, ".weight"),
+ }
+ if suffix in simple:
+ key, sfx = simple[suffix]
+ return [(tn(key, bid, sfx), data_torch)]
+
+ # kv_b_proj: split into k_b (transposed) and v_b
+ if suffix == "self_attn.kv_b_proj.weight":
+ k_b, v_b = split_kv_b_proj(data_torch, n_head, qk_nope, v_head_dim)
+ return [
+ (tn(gguf.MODEL_TENSOR.ATTN_K_B, bid), k_b),
+ (tn(gguf.MODEL_TENSOR.ATTN_V_B, bid), v_b),
+ ]
+
+ # fused stacked experts: split gate_up into gate/up
+ if suffix == "mlp.experts.gate_up_proj":
+ gate, up = split_gate_up(data_torch, moe_inter)
+ return [
+ (tn(gguf.MODEL_TENSOR.FFN_GATE_EXP, bid), gate),
+ (tn(gguf.MODEL_TENSOR.FFN_UP_EXP, bid), up),
+ ]
+ if suffix == "mlp.experts.down_proj":
+ return [(tn(gguf.MODEL_TENSOR.FFN_DOWN_EXP, bid), data_torch)]
+
+ raise ValueError(f"Unsupported HY_V4 tensor {name!r} (suffix {suffix!r})")
diff --git a/convert_hf_to_gguf_update.py b/convert_hf_to_gguf_update.py
index e5d3196efe..c4141afa61 100755
--- a/convert_hf_to_gguf_update.py
+++ b/convert_hf_to_gguf_update.py
@@ -176,6 +176,7 @@ pre_computed_hashes = [
{"name": "minerva-7b", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0", "chkhsh": "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35"},
{"name": "hunyuan", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Hunyuan-A13B-Instruct", "chkhsh": "7e57df22b1fe23a7b1e1c7f3dc4e3f96d43a4eb0836d0c6bdc3436d7b2f1c664"},
{"name": "hunyuan-dense", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Hunyuan-4B-Instruct", "chkhsh": "bba3b3366b646dbdded5dbc42d59598b849371afc42f7beafa914afaa5b70aa6"},
+ {"name": "hy_v4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Hy4-preview", "chkhsh": "e6ddf9c6686791c12d698d34c31ab9be1fea9af5a3d9a6909783ab382198ae1c"},
# falcon-h1 series uses 4 different tokenizers across model sizes (0.5b - 34b), hence we need to define 4 different hashes
{"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-0.5B-Base", "chkhsh": "a6b57017d60e6edb4d88ecc2845188e0eb333a70357e45dcc9b53964a73bbae6"},
{"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-1B-Base", "chkhsh": "60476e1243776c4fb1b993dbd7a5f15ac22f83c80afdf425fa5ae01c8d44ef86"},
diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py
index b85f62a311..399d31f1d5 100644
--- a/gguf-py/gguf/constants.py
+++ b/gguf-py/gguf/constants.py
@@ -230,6 +230,8 @@ class Keys:
COUNT = "{arch}.hyper_connection.count"
SINKHORN_ITERATIONS = "{arch}.hyper_connection.sinkhorn_iterations"
EPSILON = "{arch}.hyper_connection.epsilon"
+ # scale of the post gate (DeepSeek-V4 hardcodes 2.0)
+ MAGNITUDE = "{arch}.hyper_connection.magnitude"
# absent means the mix projection is full rank (DeepSeek-V4 behaviour)
LOW_RANK = "{arch}.hyper_connection.low_rank"
@@ -592,6 +594,7 @@ class MODEL_ARCH(IntEnum):
HUNYUAN_DENSE = auto()
HUNYUAN_VL = auto()
HY_V3 = auto()
+ HY_V4 = auto()
SMOLLM3 = auto()
GPT_OSS = auto()
LFM2 = auto()
@@ -1345,6 +1348,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.HUNYUAN_DENSE: "hunyuan-dense",
MODEL_ARCH.HUNYUAN_VL: "hunyuan_vl",
MODEL_ARCH.HY_V3: "hy_v3",
+ MODEL_ARCH.HY_V4: "hy_v4",
MODEL_ARCH.SMOLLM3: "smollm3",
MODEL_ARCH.GPT_OSS: "gpt-oss",
MODEL_ARCH.LFM2: "lfm2",
@@ -4739,6 +4743,48 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD,
MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM,
],
+ MODEL_ARCH.HY_V4: [
+ MODEL_TENSOR.TOKEN_EMBD,
+ MODEL_TENSOR.OUTPUT_NORM,
+ MODEL_TENSOR.OUTPUT,
+ MODEL_TENSOR.ROPE_FREQS,
+ MODEL_TENSOR.HC_HEAD_FN,
+ MODEL_TENSOR.HC_HEAD_BASE,
+ MODEL_TENSOR.HC_HEAD_SCALE,
+ MODEL_TENSOR.ATTN_NORM,
+ MODEL_TENSOR.ATTN_SINKS,
+ MODEL_TENSOR.ATTN_Q_A,
+ MODEL_TENSOR.ATTN_Q_A_NORM,
+ MODEL_TENSOR.ATTN_Q_B,
+ MODEL_TENSOR.ATTN_KV_A_MQA,
+ MODEL_TENSOR.ATTN_KV_A_NORM,
+ MODEL_TENSOR.ATTN_K_B,
+ MODEL_TENSOR.ATTN_V_B,
+ MODEL_TENSOR.ATTN_OUT,
+ MODEL_TENSOR.ATTN_GATE,
+ MODEL_TENSOR.INDEXER_K_NORM,
+ MODEL_TENSOR.INDEXER_PROJ,
+ MODEL_TENSOR.INDEXER_ATTN_K,
+ MODEL_TENSOR.INDEXER_ATTN_Q_B,
+ MODEL_TENSOR.HC_ATTN_FN,
+ MODEL_TENSOR.HC_ATTN_BASE,
+ MODEL_TENSOR.HC_ATTN_SCALE,
+ MODEL_TENSOR.HC_FFN_FN,
+ MODEL_TENSOR.HC_FFN_BASE,
+ MODEL_TENSOR.HC_FFN_SCALE,
+ MODEL_TENSOR.FFN_GATE_INP,
+ MODEL_TENSOR.FFN_EXP_PROBS_B,
+ MODEL_TENSOR.FFN_NORM,
+ MODEL_TENSOR.FFN_GATE,
+ MODEL_TENSOR.FFN_DOWN,
+ MODEL_TENSOR.FFN_UP,
+ MODEL_TENSOR.FFN_GATE_EXP,
+ MODEL_TENSOR.FFN_DOWN_EXP,
+ MODEL_TENSOR.FFN_UP_EXP,
+ MODEL_TENSOR.FFN_GATE_SHEXP,
+ MODEL_TENSOR.FFN_DOWN_SHEXP,
+ MODEL_TENSOR.FFN_UP_SHEXP,
+ ],
MODEL_ARCH.SMOLLM3: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
@@ -5438,6 +5484,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_ROT_EMBD,
],
+ MODEL_ARCH.HY_V4: [
+ MODEL_TENSOR.ROPE_FREQS,
+ MODEL_TENSOR.ATTN_ROT_EMBD,
+ ],
MODEL_ARCH.CHATGLM: [
MODEL_TENSOR.ROPE_FREQS,
],
diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py
index 689c2fca11..50e4d7c534 100644
--- a/gguf-py/gguf/gguf_writer.py
+++ b/gguf-py/gguf/gguf_writer.py
@@ -1055,6 +1055,9 @@ class GGUFWriter:
def add_hyper_connection_epsilon(self, value: float) -> None:
self.add_float32(Keys.HyperConnection.EPSILON.format(arch=self.arch), value)
+ def add_hyper_connection_magnitude(self, value: float) -> None:
+ self.add_float32(Keys.HyperConnection.MAGNITUDE.format(arch=self.arch), value)
+
def add_hyper_connection_low_rank(self, value: int) -> None:
self.add_uint32(Keys.HyperConnection.LOW_RANK.format(arch=self.arch), value)
diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
index 446de4ae25..d06be641a2 100644
--- a/src/llama-arch.cpp
+++ b/src/llama-arch.cpp
@@ -121,6 +121,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_HUNYUAN_DENSE, "hunyuan-dense" },
{ LLM_ARCH_HUNYUAN_VL, "hunyuan_vl" },
{ LLM_ARCH_HY_V3, "hy_v3" },
+ { LLM_ARCH_HY_V4, "hy_v4" },
{ LLM_ARCH_SMOLLM3, "smollm3" },
{ LLM_ARCH_OPENAI_MOE, "gpt-oss" },
{ LLM_ARCH_LFM2, "lfm2" },
@@ -294,6 +295,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_HYPER_CONNECTION_COUNT, "%s.hyper_connection.count" },
{ LLM_KV_HYPER_CONNECTION_SINKHORN_ITERATIONS, "%s.hyper_connection.sinkhorn_iterations" },
{ LLM_KV_HYPER_CONNECTION_EPSILON, "%s.hyper_connection.epsilon" },
+ { LLM_KV_HYPER_CONNECTION_MAGNITUDE, "%s.hyper_connection.magnitude" },
{ LLM_KV_HYPER_CONNECTION_LOW_RANK, "%s.hyper_connection.low_rank" },
{ LLM_KV_PLE_LAYERS, "%s.ple.layers" },
@@ -1130,6 +1132,7 @@ bool llm_arch_supports_sm_tensor(const llm_arch & arch) {
case LLM_ARCH_OLMOE:
case LLM_ARCH_DEEPSEEK2:
case LLM_ARCH_DEEPSEEK32:
+ case LLM_ARCH_HY_V4:
case LLM_ARCH_DOTS3NOTE:
case LLM_ARCH_GLM_DSA:
case LLM_ARCH_BITNET:
diff --git a/src/llama-arch.h b/src/llama-arch.h
index 0c0b994836..62dfa5d817 100644
--- a/src/llama-arch.h
+++ b/src/llama-arch.h
@@ -126,6 +126,7 @@ enum llm_arch {
LLM_ARCH_HUNYUAN_DENSE,
LLM_ARCH_HUNYUAN_VL,
LLM_ARCH_HY_V3,
+ LLM_ARCH_HY_V4,
LLM_ARCH_SMOLLM3,
LLM_ARCH_OPENAI_MOE,
LLM_ARCH_LFM2,
@@ -299,6 +300,7 @@ enum llm_kv {
LLM_KV_HYPER_CONNECTION_COUNT,
LLM_KV_HYPER_CONNECTION_SINKHORN_ITERATIONS,
LLM_KV_HYPER_CONNECTION_EPSILON,
+ LLM_KV_HYPER_CONNECTION_MAGNITUDE,
LLM_KV_HYPER_CONNECTION_LOW_RANK,
LLM_KV_PLE_LAYERS,
diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index 3cc27717ec..c1ef12f56b 100644
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -2317,7 +2317,8 @@ uint32_t llama_context::graph_max_nodes(uint32_t n_tokens) const {
(model.arch == LLM_ARCH_DFLASH && model.hparams.dsv4_hc_mult > 0) ||
model.arch == LLM_ARCH_NANBEIGE ||
model.arch == LLM_ARCH_MINIMAX_01 ||
- model.arch == LLM_ARCH_MINIMAX_M3) {
+ model.arch == LLM_ARCH_MINIMAX_M3 ||
+ model.arch == LLM_ARCH_HY_V4) {
res = std::max<uint32_t>(n_tokens * 40, 32u * model.n_tensors());
} else if (model.arch == LLM_ARCH_DFLASH && model.hparams.dflash_selector_rank > 0) {
// DFlash2's convolutions and selector are shape work rather than matmuls,
diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index 274a626433..8ea441f441 100644
--- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -566,7 +566,10 @@ void llm_graph_input_attn_k_dsa::set_input(const llama_ubatch * ubatch) {
mctx->get_lid()->set_input_kq_mask(self_kq_mask_lid, ubatch, cparams.causal_attn);
- mctx->get_lid()->set_input_k_rot(self_k_rot_lid);
+ // left unallocated when the indexer does not use the rotation
+ if (self_k_rot_lid && self_k_rot_lid->buffer) {
+ mctx->get_lid()->set_input_k_rot(self_k_rot_lid);
+ }
}
bool llm_graph_input_attn_k_dsa::can_reuse(const llm_graph_params & params) {
@@ -2170,7 +2173,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
const float limit = hparams.swiglu_clamp_exp[il];
constexpr float eps = 1e-6f;
if (limit > eps) {
- if (arch == LLM_ARCH_DEEPSEEK4 || (arch == LLM_ARCH_DFLASH && hparams.dsv4_hc_mult > 0)) {
+ if (arch == LLM_ARCH_DEEPSEEK4 || (arch == LLM_ARCH_DFLASH && hparams.dsv4_hc_mult > 0) || arch == LLM_ARCH_HY_V4) {
cur = ggml_swiglu_clamp(ctx0, cur, up, limit);
} else {
up = ggml_clamp(ctx0, up, -limit, limit);
diff --git a/src/llama-hparams.h b/src/llama-hparams.h
index 873399aaa4..3afa49ebe8 100644
--- a/src/llama-hparams.h
+++ b/src/llama-hparams.h
@@ -297,6 +297,9 @@ struct llama_hparams {
// 0 = full rank (DeepSeek-V4)
uint32_t hc_low_rank = 0;
+ // scale of the hyper-connection post gate (DeepSeek-V4 hardcodes 2.0)
+ float hc_magnitude = 0.0f;
+
uint32_t ple_ngram_size = 0;
uint32_t ple_heads_per_ngram = 0;
uint32_t ple_conv_kernel = 0;
diff --git a/src/llama-model-saver.cpp b/src/llama-model-saver.cpp
index 919e90eccc..df2a46d932 100644
--- a/src/llama-model-saver.cpp
+++ b/src/llama-model-saver.cpp
@@ -314,6 +314,7 @@ void llama_model_saver::add_kv_from_model() {
add_kv(LLM_KV_HYPER_CONNECTION_COUNT, hparams.dsv4_hc_mult);
add_kv(LLM_KV_HYPER_CONNECTION_SINKHORN_ITERATIONS, hparams.dsv4_hc_sinkhorn_iters);
add_kv(LLM_KV_HYPER_CONNECTION_EPSILON, hparams.dsv4_hc_eps);
+ add_kv(LLM_KV_HYPER_CONNECTION_MAGNITUDE, hparams.hc_magnitude);
add_kv(LLM_KV_HASH_LAYER_COUNT, hparams.dsv4_hash_layer_count);
add_kv(LLM_KV_HYPER_CONNECTION_LOW_RANK, hparams.hc_low_rank);
diff --git a/src/llama-model.cpp b/src/llama-model.cpp
index 9e2e0c8936..b837e27654 100644
--- a/src/llama-model.cpp
+++ b/src/llama-model.cpp
@@ -288,6 +288,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
return new llama_model_hunyuan_dense(params);
case LLM_ARCH_HY_V3:
return new llama_model_hy_v3(params);
+ case LLM_ARCH_HY_V4:
+ return new llama_model_hy_v4(params);
case LLM_ARCH_SMOLLM3:
return new llama_model_smollm3(params);
case LLM_ARCH_OPENAI_MOE:
@@ -2053,7 +2055,8 @@ void llama_model::print_info() const {
if (arch == LLM_ARCH_DEEPSEEK2 || arch == LLM_ARCH_DEEPSEEK2OCR ||
arch == LLM_ARCH_DEEPSEEK32 || arch == LLM_ARCH_GLM_DSA ||
- arch == LLM_ARCH_DOTS3NOTE || arch == LLM_ARCH_MISTRAL4) {
+ arch == LLM_ARCH_DOTS3NOTE || arch == LLM_ARCH_MISTRAL4 ||
+ arch == LLM_ARCH_HY_V4) {
LLAMA_LOG_INFO("%s: n_layer_dense_lead = %d\n", __func__, hparams.n_layer_dense_lead);
LLAMA_LOG_INFO("%s: n_lora_q = %d\n", __func__, hparams.n_lora_q);
LLAMA_LOG_INFO("%s: n_lora_kv = %d\n", __func__, hparams.n_lora_kv);
@@ -2322,6 +2325,48 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
nullptr);
}
} break;
+ case LLM_ARCH_HY_V4:
+ {
+ if (hparams.indexer_top_k == 0) {
+ // full-attention checkpoint: no indexer, so no indexer key cache
+ res = new llama_kv_cache(
+ *this,
+ hparams,
+ params.type_k,
+ params.type_v,
+ !cparams.flash_attn,
+ cparams.offload_kqv,
+ cparams.kv_unified,
+ cparams.n_ctx_seq,
+ cparams.n_seq_max,
+ 1,
+ hparams.n_swa,
+ hparams.swa_type,
+ nullptr,
+ nullptr,
+ nullptr,
+ nullptr);
+ } else {
+ // only "full" layers own an indexer, so the shared layers need no indexer cache
+ llama_kv_cache::layer_filter_cb filter_lid = [&](uint32_t il) { return hparams.is_indexer_full(il); };
+
+ res = new llama_kv_cache_dsa(
+ *this,
+ params.type_k,
+ params.type_v,
+ !cparams.flash_attn,
+ cparams.offload_kqv,
+ cparams.kv_unified,
+ cparams.n_ctx_seq,
+ cparams.n_seq_max,
+ 1,
+ hparams.n_swa,
+ hparams.swa_type,
+ nullptr,
+ filter_lid,
+ nullptr);
+ }
+ } break;
case LLM_ARCH_DOTS3NOTE:
{
GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE);
@@ -2881,6 +2926,8 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
case LLM_ARCH_DOTS3NOTE:
case LLM_ARCH_NANBEIGE:
case LLM_ARCH_POCKETTTS:
+ // HY_V4 rotates consecutive pairs, matching the reference implementation
+ case LLM_ARCH_HY_V4:
return LLAMA_ROPE_TYPE_NORM;
// the pairs of head values are offset by n_rot/2
diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
index ff926ceecd..c0c34cdd8c 100644
--- a/src/llama-vocab.cpp
+++ b/src/llama-vocab.cpp
@@ -318,6 +318,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM:
case LLAMA_VOCAB_PRE_TYPE_HUNYUAN_DENSE:
case LLAMA_VOCAB_PRE_TYPE_JOYAI_LLM:
+ case LLAMA_VOCAB_PRE_TYPE_HY_V4:
regex_exprs = {
"\\p{N}{1,3}",
"[一-龥぀-ゟ゠-ヿ]+",
@@ -2350,6 +2351,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
tokenizer_pre == "hunyuan-dense") {
pre_type = LLAMA_VOCAB_PRE_TYPE_HUNYUAN_DENSE;
clean_spaces = false;
+ } else if (
+ tokenizer_pre == "hy_v4") {
+ pre_type = LLAMA_VOCAB_PRE_TYPE_HY_V4;
+ clean_spaces = false;
} else if (
tokenizer_pre == "joyai-llm") {
pre_type = LLAMA_VOCAB_PRE_TYPE_JOYAI_LLM;
diff --git a/src/llama-vocab.h b/src/llama-vocab.h
index b7c2892633..e02ea78ffa 100644
--- a/src/llama-vocab.h
+++ b/src/llama-vocab.h
@@ -65,6 +65,7 @@ enum llama_vocab_pre_type {
LLAMA_VOCAB_PRE_TYPE_GRANITE_EMB_MULTI = 54,
LLAMA_VOCAB_PRE_TYPE_MELLUM2 = 55,
LLAMA_VOCAB_PRE_TYPE_LAGUNA = 56,
+ LLAMA_VOCAB_PRE_TYPE_HY_V4 = 57,
};
struct LLM_KV;
diff --git a/src/models/hy-v4.cpp b/src/models/hy-v4.cpp
new file mode 100644
index 0000000000..ee41787ba3
--- /dev/null
+++ b/src/models/hy-v4.cpp
@@ -0,0 +1,601 @@
+#include "models.h"
+
+#include "llama-kv-cache.h"
+#include "llama-kv-cache-dsa.h"
+
+#include <cmath>
+
+// iHC (independent Hyper-Connections) helpers. Same layout as the DeepSeek-V4 HC, but without
+// the comb/sinkhorn term: hc_fn makes only 2*hc coefficients (pre + post). The streams mix
+// through the pre-reduce / post-distribute round trip instead.
+
+static size_t hy_v4_elem_offset(const ggml_tensor * t, int64_t i) {
+ return ggml_row_size(t->type, i);
+}
+
+static ggml_tensor * hy_v4_view_1d(ggml_context * ctx, ggml_tensor * t, int64_t ne0, int64_t i0) {
+ return ggml_view_1d(ctx, t, ne0, hy_v4_elem_offset(t, i0));
+}
+
+static ggml_tensor * hy_v4_view_2d(ggml_context * ctx, ggml_tensor * t, int64_t ne0, int64_t ne1, int64_t i0) {
+ return ggml_view_2d(ctx, t, ne0, ne1, t->nb[1], hy_v4_elem_offset(t, i0));
+}
+
+void llama_model_hy_v4::load_arch_hparams(llama_model_loader & ml) {
+ ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+ ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead, false);
+ ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK, hparams.n_lora_q);
+ ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK, hparams.n_lora_kv);
+ ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA, hparams.n_embd_head_k_mla_impl);
+ ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla_impl);
+ ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
+ ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared);
+ ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale, false);
+ ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM, hparams.expert_weights_norm, false);
+ ml.get_key(LLM_KV_EXPERT_GATING_FUNC, hparams.expert_gating_func, false);
+
+ // routed-expert SwiGLU logits clamp (shared/dense experts are NOT clamped, so
+ // swiglu_clamp_shexp is intentionally left at its 0 default)
+ ml.get_key_or_arr(LLM_KV_SWIGLU_CLAMP_EXP, hparams.swiglu_clamp_exp, hparams.n_layer_all, false);
+
+ ml.get_key(LLM_KV_HYPER_CONNECTION_COUNT, hparams.dsv4_hc_mult);
+ ml.get_key(LLM_KV_HYPER_CONNECTION_EPSILON, hparams.dsv4_hc_eps);
+ ml.get_key(LLM_KV_HYPER_CONNECTION_MAGNITUDE, hparams.hc_magnitude);
+
+ // DSA is absent on the all-full_attention checkpoints, so indexer_top_k stays 0 there
+ ml.get_key(LLM_KV_ATTENTION_INDEXER_HEAD_COUNT, hparams.indexer_n_head, false);
+ ml.get_key(LLM_KV_ATTENTION_INDEXER_KEY_LENGTH, hparams.indexer_head_size, false);
+ ml.get_key(LLM_KV_ATTENTION_INDEXER_TOP_K, hparams.indexer_top_k, false);
+
+ if (hparams.indexer_top_k > 0) {
+ // the reference plumbs rms_norm_eps into the indexer k_norm LayerNorm, and build_norm
+ // reads f_norm_eps for LLM_NORM
+ hparams.f_norm_eps = hparams.f_norm_rms_eps;
+
+ if (hparams.indexer_n_head == 0 || hparams.indexer_head_size <= hparams.n_rot()) {
+ throw std::runtime_error("hy_v4: bad indexer head count / key length");
+ }
+
+ ml.get_key_or_arr(LLM_KV_ATTENTION_INDEXER_TYPES, hparams.is_indexer_full_impl, hparams.n_layer(), false);
+ if (!hparams.is_indexer_full(0)) {
+ throw std::runtime_error("hy_v4: layer 0 must own an indexer, nothing precedes it to share");
+ }
+ }
+
+ GGML_ASSERT(hparams.is_mla());
+
+ type = LLM_TYPE_UNKNOWN;
+}
+
+void llama_model_hy_v4::load_arch_tensors(llama_model_loader &) {
+ LLAMA_LOAD_LOCALS;
+
+ const int64_t n_embd_head_k_mla = hparams.n_embd_head_k_mla();
+ const int64_t n_embd_head_v_mla = hparams.n_embd_head_v_mla();
+ const int64_t n_embd_head_qk_rope = hparams.n_rot();
+ const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
+ GGML_ASSERT(n_embd_head_qk_nope >= 1);
+
+ const int64_t q_lora_rank = hparams.n_lora_q;
+ const int64_t kv_lora_rank = hparams.n_lora_kv;
+ const int64_t n_ff_exp = hparams.n_ff_exp();
+ const int64_t n_expert_shared = hparams.n_expert_shared;
+ const int64_t hc = hparams.dsv4_hc_mult;
+
+ tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+ output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+ output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0);
+
+ // global iHC head (collapses hc streams before the final norm)
+ hc_head_fn = create_tensor(tn(LLM_TENSOR_HC_HEAD_FN, "weight"), {hc * n_embd, hc}, 0);
+ hc_head_base = create_tensor(tn(LLM_TENSOR_HC_HEAD_BASE, "weight"), {hc}, 0);
+ hc_head_scale = create_tensor(tn(LLM_TENSOR_HC_HEAD_SCALE, "weight"), {1}, 0);
+
+ for (int i = 0; i < n_layer; ++i) {
+ auto & layer = layers[i];
+
+ layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+ layer.attn_sinks = create_tensor(tn(LLM_TENSOR_ATTN_SINKS, "weight", i), {n_head}, 0);
+
+ layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, 0);
+ layer.attn_q_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_A_NORM, "weight", i), {q_lora_rank}, 0);
+ layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k_mla}, 0);
+ layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + n_embd_head_qk_rope}, 0);
+ layer.attn_kv_a_norm= create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM,"weight", i), {kv_lora_rank}, 0);
+ layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K_B, "weight", i), {n_embd_head_qk_nope, kv_lora_rank, n_head}, 0);
+ layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V_B, "weight", i), {kv_lora_rank, n_embd_head_v_mla, n_head}, 0);
+ layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_embd_head_v_mla, n_embd}, 0);
+ layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), {n_embd, n_head * n_embd_head_v_mla}, 0);
+
+ // only "full" indexer layers ship weights; "shared" layers reuse their top-k
+ if (hparams.indexer_top_k > 0 && hparams.is_indexer_full(i)) {
+ const int64_t n_indexer_head = hparams.indexer_n_head;
+ const int64_t n_embd_indexer = hparams.indexer_head_size;
+
+ layer.indexer_attn_q_b = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_Q_B, "weight", i), {q_lora_rank, n_indexer_head * n_embd_indexer}, 0);
+ layer.indexer_attn_k = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_K, "weight", i), {n_embd, n_embd_indexer}, 0);
+ layer.indexer_k_norm = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM, "weight", i), {n_embd_indexer}, 0);
+ layer.indexer_k_norm_b = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM, "bias", i), {n_embd_indexer}, 0);
+ layer.indexer_proj = create_tensor(tn(LLM_TENSOR_INDEXER_PROJ, "weight", i), {n_embd, n_indexer_head}, 0);
+ }
+
+ layer.hc_attn_fn = create_tensor(tn(LLM_TENSOR_HC_ATTN_FN, "weight", i), {hc * n_embd, 2 * hc}, 0);
+ layer.hc_attn_base = create_tensor(tn(LLM_TENSOR_HC_ATTN_BASE, "weight", i), {2 * hc}, 0);
+ layer.hc_attn_scale = create_tensor(tn(LLM_TENSOR_HC_ATTN_SCALE, "weight", i), {2}, 0);
+ layer.hc_ffn_fn = create_tensor(tn(LLM_TENSOR_HC_FFN_FN, "weight", i), {hc * n_embd, 2 * hc}, 0);
+ layer.hc_ffn_base = create_tensor(tn(LLM_TENSOR_HC_FFN_BASE, "weight", i), {2 * hc}, 0);
+ layer.hc_ffn_scale = create_tensor(tn(LLM_TENSOR_HC_FFN_SCALE, "weight", i), {2}, 0);
+
+ layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+ if (i < (int) hparams.n_layer_dense_lead) {
+ layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
+ layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0);
+ layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
+ } else {
+ layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
+ layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED);
+
+ if (n_expert == 0) {
+ throw std::runtime_error("n_expert must be > 0");
+ }
+ if (n_expert_used == 0) {
+ throw std::runtime_error("n_expert_used must be > 0");
+ }
+
+ layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, 0);
+ layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, 0);
+ layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, 0);
+
+ layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
+ layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_exp * n_expert_shared, n_embd}, 0);
+ layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
+ }
+ }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_hy_v4::build_arch_graph(const llm_graph_params & params) const {
+ return std::make_unique<graph>(*this, params);
+}
+
+// reduce hc streams x[:,i,:] weighted by w[i,:] -> [n_embd, n_tokens]
+// reference runs this in fp32 (inside the float() / autocast(fp32) context)
+static ggml_tensor * hy_v4_hc_reduce(ggml_context * ctx0, ggml_tensor * x, ggml_tensor * w, int64_t hc, int64_t n_embd, int64_t nt, ggml_type out_type) {
+ ggml_tensor * x_f32 = ggml_cast(ctx0, x, GGML_TYPE_F32);
+ ggml_tensor * result = nullptr;
+ for (int64_t ih = 0; ih < hc; ++ih) {
+ ggml_tensor * xh = ggml_view_2d(ctx0, x_f32, n_embd, nt, x_f32->nb[2], ih * x_f32->nb[1]);
+ ggml_tensor * wh = ggml_view_2d(ctx0, w, 1, nt, w->nb[1], ih * w->nb[0]);
+ ggml_tensor * cur = ggml_mul(ctx0, xh, wh);
+ result = result ? ggml_add(ctx0, result, cur) : cur;
+ }
+ return ggml_cast(ctx0, result, out_type);
+}
+
+ggml_tensor * llama_model_hy_v4::graph::build_hc_pre(
+ ggml_tensor * x,
+ ggml_tensor * hc_fn,
+ ggml_tensor * hc_scale,
+ ggml_tensor * hc_base,
+ ggml_tensor ** post,
+ int il) const {
+ const int64_t hc = hparams.dsv4_hc_mult;
+ const int64_t nt = x->ne[2];
+ GGML_ASSERT(x->ne[0] == n_embd && x->ne[1] == hc);
+
+ ggml_tensor * flat = ggml_reshape_2d(ctx0, x, hc * n_embd, nt);
+ ggml_tensor * flat_norm = ggml_rms_norm(ctx0, flat, hparams.f_norm_rms_eps);
+ ggml_tensor * mixes = ggml_mul_mat(ctx0, hc_fn, flat_norm); // [2*hc, nt]
+ cb(mixes, "hc_mixes", il);
+
+ ggml_tensor * scale_pre = hy_v4_view_1d(ctx0, hc_scale, 1, 0);
+ ggml_tensor * scale_post = hy_v4_view_1d(ctx0, hc_scale, 1, 1);
+ ggml_tensor * base_pre = hy_v4_view_1d(ctx0, hc_base, hc, 0);
+ ggml_tensor * base_post = hy_v4_view_1d(ctx0, hc_base, hc, hc);
+
+ // pre = sigmoid(mixes[:hc]*scale_pre + base_pre) + eps
+ ggml_tensor * pre = hy_v4_view_2d(ctx0, mixes, hc, nt, 0);
+ pre = ggml_mul(ctx0, pre, scale_pre);
+ pre = ggml_add(ctx0, pre, base_pre);
+ pre = ggml_sigmoid(ctx0, pre);
+ pre = ggml_scale_bias(ctx0, pre, 1.0f, hparams.dsv4_hc_eps);
+ cb(pre, "hc_pre", il);
+
+ // post = magnitude*sigmoid(mixes[hc:2hc]*scale_post + base_post) + eps
+ ggml_tensor * po = hy_v4_view_2d(ctx0, mixes, hc, nt, hc);
+ po = ggml_mul(ctx0, po, scale_post);
+ po = ggml_add(ctx0, po, base_post);
+ po = ggml_sigmoid(ctx0, po);
+ po = ggml_scale(ctx0, po, hparams.hc_magnitude);
+ po = ggml_scale_bias(ctx0, po, 1.0f, hparams.dsv4_hc_eps);
+ *post = po;
+ cb(po, "hc_post_gate", il);
+
+ return hy_v4_hc_reduce(ctx0, x, pre, hc, n_embd, nt, x->type);
+}
+
+ggml_tensor * llama_model_hy_v4::graph::build_hc_post(
+ ggml_tensor * x,
+ ggml_tensor * residual,
+ ggml_tensor * post,
+ int il) const {
+ GGML_UNUSED(il);
+ const int64_t hc = hparams.dsv4_hc_mult;
+ const int64_t nt = x->ne[1];
+ GGML_ASSERT(x->ne[0] == n_embd);
+ GGML_ASSERT(residual->ne[1] == hc);
+
+ // reference HC post runs entirely in fp32 to avoid bf16 rounding accumulation
+ // across 78 layers: post.float() * x.float() + residual.float() -> .to(dtype)
+ ggml_tensor * x_f32 = ggml_cast(ctx0, x, GGML_TYPE_F32);
+ ggml_tensor * post_f32 = ggml_cast(ctx0, post, GGML_TYPE_F32);
+ ggml_tensor * res_f32 = ggml_cast(ctx0, residual, GGML_TYPE_F32);
+
+ ggml_tensor * out = nullptr;
+ for (int64_t i = 0; i < hc; ++i) {
+ ggml_tensor * res_i = ggml_view_2d(ctx0, res_f32, n_embd, nt, res_f32->nb[2], i * res_f32->nb[1]);
+ ggml_tensor * post_i = ggml_view_2d(ctx0, post_f32, 1, nt, post_f32->nb[1], i * post_f32->nb[0]);
+ ggml_tensor * cur = ggml_add(ctx0, res_i, ggml_mul(ctx0, x_f32, post_i));
+ cur = ggml_reshape_3d(ctx0, cur, n_embd, 1, nt);
+ out = out ? ggml_concat(ctx0, out, cur, 1) : cur;
+ }
+
+ // cast back to the original type (bf16)
+ out = ggml_cast(ctx0, out, residual->type);
+ return out; // [n_embd, hc, nt]
+}
+
+ggml_tensor * llama_model_hy_v4::graph::build_hc_head(
+ ggml_tensor * x,
+ ggml_tensor * hc_fn,
+ ggml_tensor * hc_scale,
+ ggml_tensor * hc_base) const {
+ const int64_t hc = hparams.dsv4_hc_mult;
+ const int64_t nt = x->ne[2];
+
+ ggml_tensor * flat = ggml_reshape_2d(ctx0, x, hc * n_embd, nt);
+ ggml_tensor * flat_norm = ggml_rms_norm(ctx0, flat, hparams.f_norm_rms_eps);
+ ggml_tensor * mixes = ggml_mul_mat(ctx0, hc_fn, flat_norm); // [hc, nt]
+ cb(mixes, "hc_head_mixes", -1);
+
+ ggml_tensor * pre = ggml_mul(ctx0, mixes, hc_scale);
+ pre = ggml_add(ctx0, pre, hc_base);
+ pre = ggml_sigmoid(ctx0, pre);
+ pre = ggml_scale_bias(ctx0, pre, 1.0f, hparams.dsv4_hc_eps);
+ cb(pre, "hc_head_pre", -1);
+
+ return hy_v4_hc_reduce(ctx0, x, pre, hc, n_embd, nt, x->type);
+}
+
+ggml_tensor * llama_model_hy_v4::graph::build_attention(
+ const llama_model & model,
+ llm_graph_input_attn_k * inp_attn,
+ ggml_tensor * cur,
+ ggml_tensor * inp_pos,
+ float kq_scale,
+ int il) const {
+ const auto & layer = model.layers[il];
+
+ const int64_t n_embd_head_k = hparams.n_embd_head_k_mla();
+ const int64_t n_embd_head_qk_rope = hparams.n_rot();
+ const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
+ const uint32_t kv_lora_rank = hparams.n_lora_kv;
+
+ ggml_tensor * q = ggml_mul_mat(ctx0, layer.wq_a, cur);
+ q = build_norm(q, layer.attn_q_a_norm, nullptr, LLM_NORM_RMS, il);
+ q = ggml_mul_mat(ctx0, layer.wq_b, q);
+
+ ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
+ ggml_row_size(q->type, n_embd_head_k), ggml_row_size(q->type, n_embd_head_k) * n_head, 0);
+ ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
+ ggml_row_size(q->type, n_embd_head_k), ggml_row_size(q->type, n_embd_head_k) * n_head,
+ ggml_row_size(q->type, n_embd_head_qk_nope));
+
+ ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, layer.wkv_a_mqa, cur);
+ ggml_tensor * kv_cmpr = ggml_view_2d(ctx0, kv_cmpr_pe, kv_lora_rank, n_tokens,
+ ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), 0);
+ ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe, n_embd_head_qk_rope, 1, n_tokens,
+ ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+ ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+ ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));
+
+ q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+ ext_factor, attn_factor, beta_fast, beta_slow);
+ cb(q_pe, "q_pe", il);
+ k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+ ext_factor, attn_factor, beta_fast, beta_slow);
+ cb(k_pe, "k_pe", il);
+
+ kv_cmpr = build_norm(kv_cmpr, layer.attn_kv_a_norm, nullptr, LLM_NORM_RMS, il);
+ cb(kv_cmpr, "kv_cmpr", il);
+
+ // MLA absorption: q_nope @ wk_b -> compressed space
+ q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3);
+ ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, layer.wk_b, q_nope);
+ q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3);
+
+ // note: rope must go first for in-place context shifting in build_rope_shift()
+ ggml_tensor * Qcur = ggml_concat(ctx0, q_nope_absorbed, q_pe, 0);
+
+ kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
+ ggml_tensor * Kcur = ggml_concat(ctx0, kv_cmpr, k_pe, 0);
+ ggml_tensor * Vcur = kv_cmpr;
+
+ // MLA-as-MQA; wo applied manually below so the gated-MLA gate can sit before o_proj
+ ggml_tensor * attn = build_attn(inp_attn,
+ nullptr, nullptr, nullptr,
+ Qcur, Kcur, Vcur, nullptr, layer.attn_sinks, layer.wv_b, kq_scale, il);
+ cb(attn, "attn_kqv", il); // [n_head * n_embd_head_v, n_tokens]
+
+ // gated MLA: elementwise sigmoid gate on the decompressed attention output
+ ggml_tensor * gate = ggml_mul_mat(ctx0, layer.wqkv_gate, cur);
+ gate = ggml_sigmoid(ctx0, gate);
+ attn = ggml_mul(ctx0, attn, gate);
+ cb(attn, "attn_gated", il);
+
+ ggml_tensor * out = build_lora_mm(layer.wo, attn);
+ cb(out, "attn_out", il);
+
+ return out;
+}
+
+ggml_tensor * llama_model_hy_v4::graph::build_indexer_top_k(
+ const llama_model & model,
+ llm_graph_input_attn_k_dsa * inp_attn_dsa,
+ ggml_tensor * cur,
+ ggml_tensor * qr,
+ ggml_tensor * inp_pos,
+ int il) const {
+ const auto & layer = model.layers[il];
+
+ const int64_t n_indexer_head = hparams.indexer_n_head;
+ const int64_t n_embd_indexer = hparams.indexer_head_size;
+ const int64_t n_embd_indexer_rope = hparams.n_rot();
+ const int64_t n_embd_indexer_nope = n_embd_indexer - n_embd_indexer_rope;
+
+ // nope rows come first, so rope only the last n_embd_indexer_rope rows, same as the MLA path
+ ggml_tensor * iq = ggml_mul_mat(ctx0, layer.indexer_attn_q_b, qr);
+
+ iq = ggml_reshape_3d(ctx0, iq, n_embd_indexer, n_indexer_head, n_tokens);
+
+ iq = ggml_rope_ext(ctx0, iq, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base,
+ freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+ iq = ggml_rope_set_offset(iq, n_embd_indexer_nope);
+ cb(iq, "indexer_q", il);
+
+ ggml_tensor * ik = ggml_mul_mat(ctx0, layer.indexer_attn_k, cur);
+
+ ik = build_norm(ik, layer.indexer_k_norm, layer.indexer_k_norm_b, LLM_NORM, il);
+
+ ik = ggml_reshape_3d(ctx0, ik, n_embd_indexer, 1, n_tokens);
+
+ ik = ggml_rope_ext(ctx0, ik, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base,
+ freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+ ik = ggml_rope_set_offset(ik, n_embd_indexer_nope);
+ cb(ik, "indexer_k", il);
+
+ // the reference applies a Hadamard rotation here, but it only helps its FP8 kernels.
+ // it is orthogonal, so it does not change q.k and we can skip it.
+
+ const auto * mctx_lid = inp_attn_dsa->mctx->get_lid();
+ const auto & k_idxs_lid = inp_attn_dsa->get_k_idxs_lid();
+ ggml_build_forward_expand(gf, mctx_lid->cpy_k(ctx0, ik, k_idxs_lid, il));
+
+ ggml_tensor * iw = ggml_mul_mat(ctx0, layer.indexer_proj, cur);
+
+ ik = mctx_lid->get_k(ctx0, il);
+
+ const auto n_stream = ik->ne[3];
+ iq = ggml_view_4d(ctx0, iq, iq->ne[0], iq->ne[1], iq->ne[2]/n_stream, n_stream,
+ iq->nb[1], iq->nb[2], iq->nb[3]/n_stream, 0);
+ iw = ggml_view_4d(ctx0, iw, iw->ne[0], iw->ne[1]/n_stream, iw->ne[2], n_stream,
+ iw->nb[1], iw->nb[2]/n_stream, iw->nb[3]/n_stream, 0);
+
+ // fold both reference scale factors into the weights before the big score tensor
+ iw = ggml_scale(ctx0, iw, 1.0f / sqrtf(float(n_embd_indexer * n_indexer_head)));
+
+ ggml_tensor * score = nullptr;
+ if (cparams.fused_lid) {
+ score = ggml_lightning_indexer(ctx0, iq, ik, iw, inp_attn_dsa->get_kq_mask_lid());
+ cb(score, "indexer_score", il);
+ res->add_fused_node({LLM_FUSED_OP_LIGHTNING_INDEXER, score, il});
+ } else {
+ iq = ggml_permute(ctx0, iq, 0, 2, 1, 3);
+ ik = ggml_permute(ctx0, ik, 0, 2, 1, 3);
+
+ score = ggml_mul_mat(ctx0, ik, iq);
+ score = ggml_cont(ctx0, ggml_permute(ctx0, score, 2, 1, 0, 3));
+ score = ggml_relu(ctx0, score);
+ score = ggml_mul(ctx0, score, iw);
+ score = ggml_sum_rows(ctx0, score);
+ score = ggml_cont(ctx0, ggml_permute(ctx0, score, 2, 1, 0, 3));
+ score = ggml_add(ctx0, score, inp_attn_dsa->get_kq_mask_lid());
+ cb(score, "indexer_score", il);
+ }
+
+ const uint32_t n_top_k = score->ne[0] < (int64_t) hparams.indexer_top_k ? score->ne[0] : hparams.indexer_top_k;
+
+ return ggml_cont(ctx0, ggml_top_k(ctx0, score, n_top_k));
+}
+
+ggml_tensor * llama_model_hy_v4::graph::build_attention_dsa(
+ const llama_model & model,
+ llm_graph_input_attn_k_dsa * inp_attn_dsa,
+ ggml_tensor * cur,
+ ggml_tensor * inp_pos,
+ ggml_tensor ** last_top_k,
+ float kq_scale,
+ int il) const {
+ const auto & layer = model.layers[il];
+
+ const int64_t n_embd_head_k = hparams.n_embd_head_k_mla();
+ const int64_t n_embd_head_qk_rope = hparams.n_rot();
+ const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
+ const uint32_t kv_lora_rank = hparams.n_lora_kv;
+
+ ggml_tensor * qr = ggml_mul_mat(ctx0, layer.wq_a, cur);
+ qr = build_norm(qr, layer.attn_q_a_norm, nullptr, LLM_NORM_RMS, il);
+
+ if (hparams.is_indexer_full(il)) {
+ *last_top_k = build_indexer_top_k(model, inp_attn_dsa, cur, qr, inp_pos, il);
+ cb(*last_top_k, "top_k", il);
+ }
+ GGML_ASSERT(*last_top_k != nullptr);
+
+ ggml_tensor * q = ggml_mul_mat(ctx0, layer.wq_b, qr);
+
+ ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
+ ggml_row_size(q->type, n_embd_head_k), ggml_row_size(q->type, n_embd_head_k) * n_head, 0);
+ ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
+ ggml_row_size(q->type, n_embd_head_k), ggml_row_size(q->type, n_embd_head_k) * n_head,
+ ggml_row_size(q->type, n_embd_head_qk_nope));
+
+ ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, layer.wkv_a_mqa, cur);
+ ggml_tensor * kv_cmpr = ggml_view_2d(ctx0, kv_cmpr_pe, kv_lora_rank, n_tokens,
+ ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), 0);
+ ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe, n_embd_head_qk_rope, 1, n_tokens,
+ ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+ ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+ ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));
+
+ q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+ ext_factor, attn_factor, beta_fast, beta_slow);
+ cb(q_pe, "q_pe", il);
+ k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+ ext_factor, attn_factor, beta_fast, beta_slow);
+ cb(k_pe, "k_pe", il);
+
+ kv_cmpr = build_norm(kv_cmpr, layer.attn_kv_a_norm, nullptr, LLM_NORM_RMS, il);
+ cb(kv_cmpr, "kv_cmpr", il);
+
+ q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3);
+ ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, layer.wk_b, q_nope);
+ q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3);
+
+ ggml_tensor * Qcur = ggml_concat(ctx0, q_nope_absorbed, q_pe, 0);
+
+ kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
+ ggml_tensor * Kcur = ggml_concat(ctx0, kv_cmpr, k_pe, 0);
+ ggml_tensor * Vcur = kv_cmpr;
+
+ ggml_tensor * attn = build_attn(inp_attn_dsa,
+ nullptr, nullptr, nullptr,
+ Qcur, Kcur, Vcur, nullptr, layer.attn_sinks, layer.wv_b, *last_top_k, kq_scale, il);
+ cb(attn, "attn_kqv", il);
+
+ ggml_tensor * gate = ggml_mul_mat(ctx0, layer.wqkv_gate, cur);
+ gate = ggml_sigmoid(ctx0, gate);
+ attn = ggml_mul(ctx0, attn, gate);
+ cb(attn, "attn_gated", il);
+
+ ggml_tensor * out = build_lora_mm(layer.wo, attn);
+ cb(out, "attn_out", il);
+
+ return out;
+}
+
+llama_model_hy_v4::graph::graph(const llama_model & model, const llm_graph_params & params) :
+ llm_graph_context(params) {
+ const int64_t hc = hparams.dsv4_hc_mult;
+ const int64_t n_embd_head_k = hparams.n_embd_head_k_mla();
+ const float kq_scale = 1.0f / sqrtf(float(n_embd_head_k));
+
+ ggml_tensor * cur;
+
+ const bool is_dsa = hparams.indexer_top_k > 0;
+
+ ggml_tensor * inp = build_inp_embd(model.tok_embd);
+ ggml_tensor * inp_pos = build_inp_pos();
+ llm_graph_input_attn_k * inp_attn = is_dsa ? nullptr : build_attn_inp_k();
+ llm_graph_input_attn_k_dsa * inp_attn_dsa = is_dsa ? build_attn_inp_k_dsa() : nullptr;
+ ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+ // top-k of the last "full" indexer layer, reused by the following "shared" layers
+ ggml_tensor * last_top_k = nullptr;
+
+ // expand the single embedding into hc parallel residual streams
+ ggml_tensor * inpL = ggml_reshape_3d(ctx0, inp, n_embd, 1, n_tokens);
+ inpL = ggml_repeat_4d(ctx0, inpL, n_embd, hc, n_tokens, 1);
+ cb(inpL, "hc_init", -1);
+
+ for (int il = 0; il < n_layer; ++il) {
+ ggml_tensor * residual = inpL;
+ ggml_tensor * post = nullptr;
+
+ cur = build_hc_pre(inpL, model.layers[il].hc_attn_fn, model.layers[il].hc_attn_scale,
+ model.layers[il].hc_attn_base, &post, il);
+ cur = build_norm(cur, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
+ cb(cur, "attn_norm", il);
+
+ cur = is_dsa
+ ? build_attention_dsa(model, inp_attn_dsa, cur, inp_pos, &last_top_k, kq_scale, il)
+ : build_attention(model, inp_attn, cur, inp_pos, kq_scale, il);
+
+ inpL = build_hc_post(cur, residual, post, il);
+ cb(inpL, "hc_attn_out", il);
+
+ residual = inpL;
+ cur = build_hc_pre(inpL, model.layers[il].hc_ffn_fn, model.layers[il].hc_ffn_scale,
+ model.layers[il].hc_ffn_base, &post, il);
+ cur = build_norm(cur, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il);
+ cb(cur, "ffn_norm", il);
+
+ const auto & layer = model.layers[il];
+ if ((uint32_t) il < hparams.n_layer_dense_lead) {
+ cur = build_ffn(cur,
+ layer.ffn_up, NULL, NULL,
+ layer.ffn_gate, NULL, NULL,
+ layer.ffn_down, NULL, NULL,
+ NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+ cb(cur, "ffn_out", il);
+ } else {
+ ggml_tensor * moe_out = build_moe_ffn(cur,
+ layer.ffn_gate_inp,
+ layer.ffn_up_exps,
+ layer.ffn_gate_exps,
+ layer.ffn_down_exps,
+ layer.ffn_exp_probs_b,
+ n_expert, n_expert_used,
+ LLM_FFN_SILU, hparams.expert_weights_norm,
+ hparams.expert_weights_scale,
+ (llama_expert_gating_func_type) hparams.expert_gating_func,
+ il,
+ nullptr,
+ nullptr);
+ cb(moe_out, "ffn_moe_out", il);
+
+ ggml_tensor * ffn_shexp = build_ffn(cur,
+ layer.ffn_up_shexp, NULL, NULL,
+ layer.ffn_gate_shexp, NULL, NULL,
+ layer.ffn_down_shexp, NULL, NULL,
+ NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+ cb(ffn_shexp, "ffn_shexp", il);
+
+ cur = ggml_add(ctx0, moe_out, ffn_shexp);
+ cb(cur, "ffn_out", il);
+ }
+
+ inpL = build_hc_post(cur, residual, post, il);
+ cb(inpL, "l_out", il);
+ }
+
+ // prune to the requested output rows once, after all HC streams are done
+ if (inp_out_ids) {
+ ggml_tensor * flat = ggml_reshape_2d(ctx0, inpL, n_embd * hc, n_tokens);
+ flat = ggml_get_rows(ctx0, flat, inp_out_ids);
+ inpL = ggml_reshape_3d(ctx0, flat, n_embd, hc, n_outputs);
+ }
+
+ cur = build_hc_head(inpL, model.hc_head_fn, model.hc_head_scale, model.hc_head_base);
+ cb(cur, "hc_head", -1);
+
+ cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);
+ cb(cur, "result_norm", -1);
+ res->t_embd = cur;
+
+ cur = ggml_mul_mat(ctx0, model.output, cur);
+ cb(cur, "result_output", -1);
+ res->t_logits = cur;
+
+ ggml_build_forward_expand(gf, cur);
+}
diff --git a/src/models/models.h b/src/models/models.h
index 9b87a40d5a..93a6b34945 100644
--- a/src/models/models.h
+++ b/src/models/models.h
@@ -1981,6 +1981,69 @@ struct llama_model_hy_v3 : public llama_model_base {
};
+struct llama_model_hy_v4 : public llama_model_base {
+ llama_model_hy_v4(const struct llama_model_params & params) : llama_model_base(params) {}
+ void load_arch_hparams(llama_model_loader & ml) override;
+ void load_arch_tensors(llama_model_loader & ml) override;
+
+ struct graph : public llm_graph_context {
+ graph(const llama_model & model, const llm_graph_params & params);
+
+ // iHC (independent Hyper-Connections): pre reduces the hc streams to one and returns the
+ // per-stream post gates, post writes the sublayer output back into the streams, head
+ // collapses the streams before the final norm.
+ ggml_tensor * build_hc_pre(
+ ggml_tensor * x,
+ ggml_tensor * hc_fn,
+ ggml_tensor * hc_scale,
+ ggml_tensor * hc_base,
+ ggml_tensor ** post,
+ int il) const;
+
+ ggml_tensor * build_hc_post(
+ ggml_tensor * x,
+ ggml_tensor * residual,
+ ggml_tensor * post,
+ int il) const;
+
+ ggml_tensor * build_hc_head(
+ ggml_tensor * x,
+ ggml_tensor * hc_fn,
+ ggml_tensor * hc_scale,
+ ggml_tensor * hc_base) const;
+
+ ggml_tensor * build_attention(
+ const llama_model & model,
+ llm_graph_input_attn_k * inp_attn,
+ ggml_tensor * cur,
+ ggml_tensor * inp_pos,
+ float kq_scale,
+ int il) const;
+
+ // DSA lightning indexer: top-k KV positions for this layer. Only "full" layers compute
+ // it, "shared" layers reuse the last preceding full layer result through last_top_k.
+ ggml_tensor * build_indexer_top_k(
+ const llama_model & model,
+ llm_graph_input_attn_k_dsa * inp_attn_dsa,
+ ggml_tensor * cur,
+ ggml_tensor * qr,
+ ggml_tensor * inp_pos,
+ int il) const;
+
+ ggml_tensor * build_attention_dsa(
+ const llama_model & model,
+ llm_graph_input_attn_k_dsa * inp_attn_dsa,
+ ggml_tensor * cur,
+ ggml_tensor * inp_pos,
+ ggml_tensor ** last_top_k,
+ float kq_scale,
+ int il) const;
+ };
+
+ std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
+};
+
+
struct llama_model_hunyuan_vl : public llama_model_base {
llama_model_hunyuan_vl(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
diff --git a/tests/test-llama-archs.cpp b/tests/test-llama-archs.cpp
index b2ea245ab8..0f3d1c79a7 100644
--- a/tests/test-llama-archs.cpp
+++ b/tests/test-llama-archs.cpp
@@ -118,7 +118,8 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
|| arch == LLM_ARCH_KIMI_LINEAR
|| arch == LLM_ARCH_BAILINGMOE3
|| arch == LLM_ARCH_KIMI_K3
- || arch == LLM_ARCH_MISTRAL4) {
+ || arch == LLM_ARCH_MISTRAL4
+ || arch == LLM_ARCH_HY_V4) {
n_embd = 128;
n_head = 1;
n_ff = 192;
@@ -191,7 +192,8 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
|| arch == LLM_ARCH_KIMI_LINEAR
|| arch == LLM_ARCH_BAILINGMOE3
|| arch == LLM_ARCH_KIMI_K3
- || arch == LLM_ARCH_MISTRAL4) {
+ || arch == LLM_ARCH_MISTRAL4
+ || arch == LLM_ARCH_HY_V4) {
ms.add_kv(LLM_KV_ATTENTION_KEY_LENGTH, uint32_t(576));
ms.add_kv(LLM_KV_ATTENTION_VALUE_LENGTH, uint32_t(512));
ms.add_kv(LLM_KV_ROPE_DIMENSION_COUNT, uint32_t(64));
@@ -291,6 +293,22 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
ms.add_kv(LLM_KV_ATTENTION_INDEXER_LOCAL_BLOCKS, uint32_t(1));
ms.add_kv(LLM_KV_ROPE_DIMENSION_SECTIONS, std::vector<uint32_t>({n_embd_head/4, n_embd_head/4, n_embd_head/4, n_embd_head/4}));
+ if (arch == LLM_ARCH_HY_V4) {
+ ms.add_kv(LLM_KV_HYPER_CONNECTION_COUNT, uint32_t(4));
+ ms.add_kv(LLM_KV_HYPER_CONNECTION_EPSILON, 1.0e-6f);
+ ms.add_kv(LLM_KV_HYPER_CONNECTION_MAGNITUDE, 2.0f);
+ ms.add_kv(LLM_KV_SWIGLU_CLAMP_EXP, 10.0f);
+ ms.add_kv(LLM_KV_EXPERT_WEIGHTS_SCALE, 1.0f);
+ ms.add_kv(LLM_KV_EXPERT_WEIGHTS_NORM, true);
+ // layer 0 must own an indexer, the odd layers share it
+ std::vector<uint32_t> indexer_types;
+ indexer_types.reserve(n_layer);
+ for (uint32_t il = 0; il < n_layer; il++) {
+ indexer_types.push_back(il % 2 ? 0 : 1);
+ }
+ ms.add_kv(LLM_KV_ATTENTION_INDEXER_TYPES, indexer_types);
+ }
+
if (arch == LLM_ARCH_DEEPSEEK4) {
ms.add_kv(LLM_KV_ATTENTION_OUTPUT_GROUP_COUNT, uint32_t(8));
ms.add_kv(LLM_KV_ATTENTION_OUTPUT_LORA_RANK, uint32_t(32));
@@ -468,6 +486,7 @@ static bool moe_mandatory(const llm_arch arch) {
case LLM_ARCH_ERNIE4_5_MOE:
case LLM_ARCH_HUNYUAN_MOE:
case LLM_ARCH_HY_V3:
+ case LLM_ARCH_HY_V4:
case LLM_ARCH_OPENAI_MOE:
case LLM_ARCH_LFM2MOE:
case LLM_ARCH_SMALLTHINKER: