Add MiniMax-M2, Mistral4, and GLM-DSA GGUF graphs

src/mobius/__main__.py

@@ -180,6 +180,14 @@ def _resolve_static_cache_task(model_type: str) -> ModelTask:
                 "--static-cache cannot represent Falcon-H1's per-layer K, V, "
                 "convolution, and SSM states"
             )
+        if model_type == "glm_moe_dsa":
+            raise ValueError(
+                "--static-cache cannot represent GLM-DSA's per-layer-varying packed DSA cache"
+            )
+        if model_type == "mistral4_gguf":
+            raise ValueError(
+                "--static-cache is not implemented for Mistral4's latent K-only cache"
+            )
         if model_type == "gemma4":
             from mobius.tasks._gemma4 import Gemma4Task
 

src/mobius/_registry.py

@@ -111,6 +111,8 @@
     Llama4CausalLMModel,
     MageVLForConditionalGeneration,
     MaincoderCausalLMModel,
+    MiniMaxM2GGUFCausalLMModel,
+    Mistral4GGUFCausalLMModel,
     MoECausalLMModel,
     MoonshineForConditionalGeneration,
     NanoChatCausalLMModel,
@@ -492,6 +494,17 @@ def _detect_fallback_registration(hf_config) -> ModelRegistration | None:
         test_model_id="Maincode/Maincoder-1B",
         test_revision="088ec98640bdeb105f46a9ef6a1370ed5d0d2ea5",
     ),
+    "minimax_m2_gguf": ModelRegistration(
+        MiniMaxM2GGUFCausalLMModel,
+        family="minimax",
+        variant="gguf",
+    ),
+    "mistral4_gguf": ModelRegistration(
+        Mistral4GGUFCausalLMModel,
+        task="mistral4-gguf-text-generation",
+        family="mistral",
+        variant="gguf",
+    ),
     "bloom": ModelRegistration(BloomCausalLMModel),
     "orion": ModelRegistration(LayerNormCausalLMModel),
     "chatglm": ModelRegistration(ChatGLMCausalLMModel),

src/mobius/components/_deepseek_mla.py

@@ -45,6 +45,7 @@ def __init__(
         config: ArchitectureConfig,
         scale: float | None = None,
         linear_class: type | None = None,
+        split_kv_b: bool = False,
     ):
         super().__init__()
         if linear_class is None:
@@ -82,12 +83,25 @@ def __init__(
             bias=False,
         )
         self.kv_a_layernorm = RMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps)
-        # Decompresses latent KV into per-head k_nope + v
-        self.kv_b_proj = linear_class(
-            self.kv_lora_rank,
-            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
-            bias=False,
-        )
+        self._split_kv_b = split_kv_b
+        if split_kv_b:
+            self.k_b_proj = linear_class(
+                self.kv_lora_rank,
+                self.num_heads * self.qk_nope_head_dim,
+                bias=False,
+            )
+            self.v_b_proj = linear_class(
+                self.kv_lora_rank,
+                self.num_heads * self.v_head_dim,
+                bias=False,
+            )
+        else:
+            # Decompresses latent KV into per-head k_nope + v
+            self.kv_b_proj = linear_class(
+                self.kv_lora_rank,
+                self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
+                bias=False,
+            )
 
         self.o_proj = linear_class(
             self.num_heads * self.v_head_dim,
@@ -154,18 +168,28 @@ def forward(
 
         # Decompress latent KV → per-head k_nope + v
         k_pass = self.kv_a_layernorm(op, k_pass)
-        kv_decompressed = self.kv_b_proj(op, k_pass)
-        # (B, S, num_heads * (nope + v_dim)) → (B, S, num_heads, nope + v_dim)
-        kv_decompressed = op.Reshape(
-            kv_decompressed,
-            [0, 0, self.num_heads, self.qk_nope_head_dim + self.v_head_dim],
-        )
-        k_nope, value_states = op.Split(
-            kv_decompressed,
-            [self.qk_nope_head_dim, self.v_head_dim],
-            axis=-1,
-            _outputs=2,
-        )
+        if self._split_kv_b:
+            k_nope = op.Reshape(
+                self.k_b_proj(op, k_pass),
+                [0, 0, self.num_heads, self.qk_nope_head_dim],
+            )
+            value_states = op.Reshape(
+                self.v_b_proj(op, k_pass),
+                [0, 0, self.num_heads, self.v_head_dim],
+            )
+        else:
+            kv_decompressed = self.kv_b_proj(op, k_pass)
+            # (B, S, num_heads * (nope + v_dim)) → (B, S, num_heads, nope + v_dim)
+            kv_decompressed = op.Reshape(
+                kv_decompressed,
+                [0, 0, self.num_heads, self.qk_nope_head_dim + self.v_head_dim],
+            )
+            k_nope, value_states = op.Split(
+                kv_decompressed,
+                [self.qk_nope_head_dim, self.v_head_dim],
+                axis=-1,
+                _outputs=2,
+            )
         # k_nope: (B, S, H, nope_dim) → (B, S, H*nope_dim)... not needed yet
         # value_states: (B, S, H, v_dim) → (B, S, H*v_dim) for Attention op
         value_states = op.Reshape(value_states, [0, 0, -1])

src/mobius/integrations/gguf/_arch_registry.py

@@ -505,11 +505,6 @@
         "Q/K permutation, optional long/short RoPE tensors, and a conditional dense-or-MoE "
         "loader. The existing MiniCPM graph does not prove this complete GGUF contract."
     ),
-    "minimax-m2": (
-        "MiniMax-M2 uses full-vector Q/K norms, partial RoPE, and all-layer "
-        "correction-biased routed experts under metadata-selected gating. Mobius has no "
-        "exact graph or suffix-safe expert import for that topology."
-    ),
     "minimax-m3": (
         "MiniMax-M3 adds F32 sparse-indexer tensors and a second index-key cache with "
         "position/cell maps, block masks, rollback, and reorder semantics alongside main "
@@ -584,11 +579,6 @@
         "KV cache, and three chained heads selected by offsets. Mobius permits one head "
         "and cannot preserve that state or FP8 converter transform."
     ),
-    "mistral4": (
-        "Mistral4 has no NextN metadata or MTP graph; it inherits Mistral3's conditional "
-        "dense/MoE tensor loader and overrides graph construction. It is not llama, "
-        "mistral alias text, or any DeepSeek/Qwen MTP family."
-    ),
     "step35": (
         "Step3.5 executes one or more interleaved-SWA NextN heads with optional gates, "
         "routed/shared experts, centered-norm transforms, per-layer head geometry, and "
@@ -1334,6 +1324,42 @@
             "runtime packaging remains tracked by #605."
         ),
     ),
+    GGUFArchitectureSpec(
+        gguf_arch="minimax-m2",
+        model_type="minimax_m2_gguf",
+        module_type="minimax_m2_gguf",
+        config_key_map="minimax_m2",
+        config_postprocessor="minimax_m2",
+        tensor_map_recipe=("minimax_m2",),
+        required_metadata=(
+            "context_length",
+            "embedding_length",
+            "feed_forward_length",
+            "block_count",
+            "attention.head_count",
+            "attention.head_count_kv",
+            "attention.key_length",
+            "attention.value_length",
+            "attention.layer_norm_rms_epsilon",
+            "rope.freq_base",
+            "rope.dimension_count",
+            "expert_count",
+            "expert_used_count",
+            "expert_feed_forward_length",
+            "expert_gating_func",
+        ),
+        runtime=Support.DEFERRED,
+        quantized_import=Support.REJECTED,
+        reason=(
+            "Exact explicit-float import owns MiniMax-M2's non-square Q/O geometry, "
+            "full-vector Q/K RMSNorm, partial NeoX RoPE, standard dynamic/static KV cache, "
+            "and F32 selection-biased sigmoid MoE routing. Loader-accepted Q/K/V biases "
+            "and fused QKV are rejected because the pinned graph ignores or cannot execute "
+            "them. Packed projection/expert import and runtime packaging remain deferred; "
+            "the smallest immutable public GGUF is 46,514,882,176 bytes, above the "
+            "16 GiB evidence budget. Use keep_quantized=False."
+        ),
+    ),
     GGUFArchitectureSpec(
         gguf_arch="plamo",
         model_type="plamo",
@@ -2131,6 +2157,48 @@
         runtime=Support.DEFERRED,
         reason=_DEEPSEEK4_GGUF_GRAPH_REASON,
     ),
+    GGUFArchitectureSpec(
+        gguf_arch="mistral4",
+        model_type="mistral4_gguf",
+        module_type="mistral4_gguf",
+        config_key_map="mistral4",
+        config_postprocessor="mistral4",
+        tensor_map_recipe=("mistral4",),
+        tensor_processor="mistral4",
+        required_metadata=(
+            "context_length",
+            "embedding_length",
+            "feed_forward_length",
+            "block_count",
+            "attention.head_count",
+            "attention.head_count_kv",
+            "attention.key_length",
+            "attention.value_length",
+            "attention.key_length_mla",
+            "attention.value_length_mla",
+            "attention.q_lora_rank",
+            "attention.kv_lora_rank",
+            "attention.layer_norm_rms_epsilon",
+            "rope.freq_base",
+            "rope.dimension_count",
+            "expert_count",
+            "expert_used_count",
+            "expert_feed_forward_length",
+            "expert_shared_count",
+        ),
+        rope_interleave=True,
+        runtime=Support.DEFERRED,
+        quantized_import=Support.REJECTED,
+        reason=(
+            "Exact explicit-float import owns Mistral4's DeepSeek-V2 MLA projections, "
+            "dense-prefix plus mandatory shared/routed MoE blocks, YaRN scaling, and one "
+            "graph-visible latent-plus-RoPE K cache per layer. Legacy non-MLA, Q-LoRA-free, "
+            "temperature-scaled, and NextN layouts fail closed. Packed import and runtime "
+            "packaging remain deferred; the smallest immutable public GGUF is "
+            "32,306,941,632 bytes, above the 16 GiB evidence budget. "
+            "Use keep_quantized=False."
+        ),
+    ),
     # GLM-5.2 GGUFs (e.g. unsloth/GLM-5.2-GGUF) tag the architecture 'glm-dsa'
     # (MLA + DeepSeek Sparse Attention + MoE) and mobius's registry key is
     # 'glm_moe_dsa'. The format bridge is keyed on the authoritative
@@ -2142,12 +2210,43 @@
         model_type="glm_moe_dsa",
         aliases=frozenset({"glm_dsa"}),
         config_key_map="glm_dsa",
-        tensor_map=Support.DEFERRED,
+        config_postprocessor="glm_dsa",
+        tensor_map_recipe=("glm_dsa",),
+        tensor_processor="glm_dsa",
+        required_metadata=(
+            "context_length",
+            "embedding_length",
+            "feed_forward_length",
+            "block_count",
+            "attention.head_count",
+            "attention.head_count_kv",
+            "attention.key_length",
+            "attention.value_length",
+            "attention.key_length_mla",
+            "attention.value_length_mla",
+            "attention.q_lora_rank",
+            "attention.kv_lora_rank",
+            "attention.layer_norm_rms_epsilon",
+            "rope.freq_base",
+            "rope.dimension_count",
+            "expert_count",
+            "expert_used_count",
+            "expert_feed_forward_length",
+            "expert_shared_count",
+            "attention.indexer.head_count",
+            "attention.indexer.key_length",
+            "attention.indexer.top_k",
+        ),
+        rope_interleave=True,
+        quantized_import=Support.REJECTED,
         reason=(
-            "Config extraction and the glm_moe_dsa graph are both available, but "
-            "no GGUF→HuggingFace tensor-name mapping has been written for GLM-5.2's "
-            "MLA + DSA-indexer tensor families yet, so weights cannot be routed "
-            "into the graph. " + _NO_TENSOR_MAP
+            "Exact explicit-float MLA, DSA indexer, dense-prefix/routed-MoE tensor routing, "
+            "and packed dynamic-cache graph construction are covered. Packed import remains "
+            "rejected because GLM-5.2's rank-3 K/V-B transforms and correction-biased routed "
+            "experts do not have an independently proven storage-preserving runtime path. "
+            "The released GGUF also carries a routed DSA/MLA MTP block outside the current "
+            "sidecar ABI, and representative real-weight runtime evidence exceeds the 16 GiB "
+            "artifact budget. Use keep_quantized=False."
         ),
     ),
     GGUFArchitectureSpec(

src/mobius/integrations/gguf/_arch_registry_test.py

@@ -60,7 +60,7 @@
 #: Number of importable architectures. Pinned so that adding support is a
 #: deliberate act that also updates the documented support matrix, and so that
 #: accidentally losing an architecture is a failure rather than a silence.
-_EXPECTED_SUPPORTED_COUNT = 102
+_EXPECTED_SUPPORTED_COUNT = 105
 _PROMOTED_CONVENTIONAL_DECODERS = frozenset(
     {
         "bitnet",
@@ -85,6 +85,8 @@
         "gemma-embedding",
         "llama-embed",
         "maincoder",
+        "minimax-m2",
+        "mistral4",
         "starcoder",
         "xverse",
     }
@@ -328,6 +330,8 @@ def test_every_float_importable_architecture_has_a_quantized_verdict(self) -> No
             "mamba",
             "mamba2",
             "maincoder",
+            "minimax-m2",
+            "mistral4",
             "nemotron_h",
             "nemotron_h_moe",
             "neo-bert",
@@ -343,6 +347,7 @@ def test_every_float_importable_architecture_has_a_quantized_verdict(self) -> No
             "grok",
             "grovemoe",
             "hunyuan-moe",
+            "glm-dsa",
             "hy_v3",
             "jais",
             "mpt",