Enable direct GLM-5.2 GGUF tensor mapping

docs/api/build_from_gguf.md

@@ -9,7 +9,7 @@ Support is capability-specific: graph import does not imply runtime packaging.
 
 | Census | Total | Closure |
 |---|---:|---|
-| Architectures | 148 | graph verdicts: {'deferred': 46, 'rejected': 2, 'supported': 100}; importable: 99; quantized import: {'rejected': 41, 'supported': 107}; runtime: {'deferred': 142, 'rejected': 2, 'supported': 4} |
+| Architectures | 148 | graph verdicts: {'deferred': 46, 'rejected': 2, 'supported': 100}; importable: 100; quantized import: {'rejected': 41, 'supported': 107}; runtime: {'deferred': 142, 'rejected': 2, 'supported': 4} |
 | Active stored qtypes | 25 | 24 have an import route; 1 are explicitly deferred with no route |
 | Serialized projector strings | 60 | {'graph-importable': 9, 'runtime-supported': 0} |
 | Tokenizer pre identifiers | 87 | 56 semantic groups; route dispositions: {'deferred-compiled-semantics': 45, 'deferred-pinned-artifact-evidence': 11, 'validated-pinned-source': 31} |
@@ -98,13 +98,12 @@ remain machine-readable in `_route_census.py`; this table groups only shared nex
 | `dependency-or-runtime-abi-blocked` | `mtp-specialized-abi` | `mtp:bailingmoe3`, `mtp:cohere2moe`, `mtp:deepseek2`, `mtp:deepseek32`, `mtp:deepseek4`, `mtp:glm-dsa`, `mtp:mimo2`, `mtp:nemotron_h_moe`, `mtp:qwen35moe`, `mtp:qwen3next`, `mtp:step35` | specialized sidecar graph; routed/cache state ABI |
 | `dependency-or-runtime-abi-blocked` | `projector-runtime-abi` | `projector:resampler` | dynamic processor-to-graph media shape ABI |
 | `dependency-or-runtime-abi-blocked` | `tokenizer-compiled-semantics` | `tokenizer:afmoe`, `tokenizer:bloom`, `tokenizer:chameleon`, `tokenizer:codeshell`, `tokenizer:command-r`, `tokenizer:dbrx`, `tokenizer:deepseek-coder`, `tokenizer:deepseek-llm`, `tokenizer:deepseek-v3`, `tokenizer:default`, `tokenizer:exaone`, `tokenizer:exaone-moe`, `tokenizer:falcon`, `tokenizer:gpt3-finnish`, `tokenizer:granite-docling`, `tokenizer:granite-embed-multi-97m`, `tokenizer:grok-2`, `tokenizer:hunyuan`, `tokenizer:hunyuan-dense`, `tokenizer:jais`, `tokenizer:jais-2`, `tokenizer:joyai-llm`, `tokenizer:kimi-k2`, `tokenizer:laguna`, `tokenizer:megrez`, `tokenizer:mellum2`, `tokenizer:minerva-7b`, `tokenizer:minicpm5`, `tokenizer:minimax-m2`, `tokenizer:mpt`, `tokenizer:olmo`, `tokenizer:poro-chat`, `tokenizer:refact`, `tokenizer:sarvam-moe`, `tokenizer:seed-coder`, `tokenizer:smaug-bpe`, `tokenizer:solar-open`, `tokenizer:stablelm2`, `tokenizer:starcoder`, `tokenizer:superbpe`, `tokenizer:tekken`, `tokenizer:trillion`, `tokenizer:viking`, `tokenizer:whitespace`, `tokenizer:youtu` | compiled pinned llama.cpp oracle; dispatch-equivalence fixture |
-| `evidence-only` | `architecture-runtime-evidence` | `architecture:apertus`, `architecture:arcee`, `architecture:arctic`, `architecture:baichuan`, `architecture:bailingmoe`, `architecture:bert`, `architecture:bitnet`, `architecture:bloom`, `architecture:chatglm`, `architecture:codeshell`, `architecture:cohere2`, `architecture:command-r`, `architecture:dbrx`, `architecture:deci`, `architecture:deepseek`, `architecture:dflash`, `architecture:dots1`, `architecture:dream`, `architecture:eagle3`, `architecture:ernie4_5`, `architecture:ernie4_5-moe`, `architecture:eurobert`, `architecture:exaone`, `architecture:falcon`, `architecture:gemma`, `architecture:gemma-embedding`, `architecture:gemma2`, `architecture:gemma3`, `architecture:gemma4`, `architecture:gpt2`, `architecture:gptneox`, `architecture:granite`, `architecture:granitemoe`, `architecture:hunyuan-dense`, `architecture:hy_v3`, `architecture:internlm2`, `architecture:jais`, `architecture:jais2`, `architecture:jina-bert-v2`, `architecture:jina-bert-v3`, `architecture:lfm2moe`, `architecture:llada`, `architecture:llada-moe`, `architecture:llama-embed`, `architecture:maincoder`, `architecture:mamba`, `architecture:mamba2`, `architecture:minicpm`, `architecture:minicpm3`, `architecture:modern-bert`, `architecture:mpt`, `architecture:muse-glimmer`, `architecture:nemotron`, `architecture:nemotron_h`, `architecture:neo-bert`, `architecture:nomic-bert`, `architecture:nomic-bert-moe`, `architecture:olmo`, `architecture:olmo2`, `architecture:olmoe`, `architecture:openelm`, `architecture:orion`, `architecture:pangu-embedded`, `architecture:phi2`, `architecture:phi3`, `architecture:phimoe`, `architecture:plamo`, `architecture:plm`, `architecture:qwen`, `architecture:qwen2moe`, `architecture:qwen2vl`, `architecture:qwen3`, `architecture:qwen35`, `architecture:qwen3moe`, `architecture:qwen3next`, `architecture:refact`, `architecture:rnd1`, `architecture:seed_oss`, `architecture:smallthinker`, `architecture:smollm3`, `architecture:stablelm`, `architecture:starcoder`, `architecture:starcoder2`, `architecture:t5`, `architecture:t5encoder`, `architecture:talkie`, `architecture:xverse` | immutable representative GGUF; full-logit prefill and cached-decode parity; deterministic generation/state evidence |
+| `evidence-only` | `architecture-runtime-evidence` | `architecture:apertus`, `architecture:arcee`, `architecture:arctic`, `architecture:baichuan`, `architecture:bailingmoe`, `architecture:bert`, `architecture:bitnet`, `architecture:bloom`, `architecture:chatglm`, `architecture:codeshell`, `architecture:cohere2`, `architecture:command-r`, `architecture:dbrx`, `architecture:deci`, `architecture:deepseek`, `architecture:dflash`, `architecture:dots1`, `architecture:dream`, `architecture:eagle3`, `architecture:ernie4_5`, `architecture:ernie4_5-moe`, `architecture:eurobert`, `architecture:exaone`, `architecture:falcon`, `architecture:gemma`, `architecture:gemma-embedding`, `architecture:gemma2`, `architecture:gemma3`, `architecture:gemma4`, `architecture:glm-dsa`, `architecture:gpt2`, `architecture:gptneox`, `architecture:granite`, `architecture:granitemoe`, `architecture:hunyuan-dense`, `architecture:hy_v3`, `architecture:internlm2`, `architecture:jais`, `architecture:jais2`, `architecture:jina-bert-v2`, `architecture:jina-bert-v3`, `architecture:lfm2moe`, `architecture:llada`, `architecture:llada-moe`, `architecture:llama-embed`, `architecture:maincoder`, `architecture:mamba`, `architecture:mamba2`, `architecture:minicpm`, `architecture:minicpm3`, `architecture:modern-bert`, `architecture:mpt`, `architecture:muse-glimmer`, `architecture:nemotron`, `architecture:nemotron_h`, `architecture:neo-bert`, `architecture:nomic-bert`, `architecture:nomic-bert-moe`, `architecture:olmo`, `architecture:olmo2`, `architecture:olmoe`, `architecture:openelm`, `architecture:orion`, `architecture:pangu-embedded`, `architecture:phi2`, `architecture:phi3`, `architecture:phimoe`, `architecture:plamo`, `architecture:plm`, `architecture:qwen`, `architecture:qwen2moe`, `architecture:qwen2vl`, `architecture:qwen3`, `architecture:qwen35`, `architecture:qwen3moe`, `architecture:qwen3next`, `architecture:refact`, `architecture:rnd1`, `architecture:seed_oss`, `architecture:smallthinker`, `architecture:smollm3`, `architecture:stablelm`, `architecture:starcoder`, `architecture:starcoder2`, `architecture:t5`, `architecture:t5encoder`, `architecture:talkie`, `architecture:xverse` | immutable representative GGUF; full-logit prefill and cached-decode parity; deterministic generation/state evidence |
 | `evidence-only` | `draft-runtime-evidence` | `draft:dflash`, `draft:eagle3` | target acceptance loop; draft cache orchestration; deterministic speedup parity |
 | `evidence-only` | `mtp-runtime-evidence` | `mtp:hy_v3`, `mtp:qwen35` | target acceptance loop; cache-threaded draft/target parity |
 | `evidence-only` | `projector-runtime-evidence` | `projector:adapter`, `projector:gemma3`, `projector:gemma4v`, `projector:ldp`, `projector:ldpv2`, `projector:mlp`, `projector:muse-glimmer`, `projector:qwen2.5vl_merger`, `projector:qwen2vl_merger` | paired text target; processor boundary; deterministic multimodal package execution |
 | `evidence-only` | `tokenizer-artifact-evidence` | `tokenizer:bailingmoe`, `tokenizer:bailingmoe2`, `tokenizer:chatglm-bpe`, `tokenizer:cohere2moe`, `tokenizer:glm4`, `tokenizer:llada-moe`, `tokenizer:tiny_aya` | immutable GGUF/source pair; ordered vocabulary and encoding parity |
 | `immediately-implementable` | `architecture-implementation` | `architecture:grok`, `architecture:grovemoe`, `architecture:hunyuan-moe`, `architecture:minimax-m2`, `architecture:mistral4` | exact metadata extraction; tensor closure; dedicated graph and parity |
-| `immediately-implementable` | `architecture-implementation` | `architecture:glm-dsa` | suffix-exact tensor mapping; packed-value transform proof; graph closure |
 | `immediately-implementable` | `projector-implementation` | `projector:cogvlm`, `projector:deepseekocr`, `projector:deepseekocr2`, `projector:dots3note_a`, `projector:dots3note_v`, `projector:dots_ocr`, `projector:exaone4_5`, `projector:gemma3na`, `projector:gemma3nv`, `projector:gemma4a`, `projector:gemma4ua`, `projector:gemma4uv`, `projector:glm4v`, `projector:glma`, `projector:granite4_vision`, `projector:granite_speech`, `projector:hunyuanvl`, `projector:idefics3`, `projector:internvl`, `projector:janus_pro`, `projector:kimik25`, `projector:kimivl`, `projector:lfm2`, `projector:lfm2a`, `projector:lightonocr`, `projector:llama4`, `projector:meralion`, `projector:mimo_audio`, `projector:mimovl`, `projector:minicpmv4_6`, `projector:minimax_m3`, `projector:musicflamingo`, `projector:nemotron_v2_vl`, `projector:paddleocr`, `projector:parakeet`, `projector:pixtral`, `projector:pockettts_spkenc`, `projector:qwen2.5o`, `projector:qwen2a`, `projector:qwen3a`, `projector:qwen3tts_spkenc`, `projector:qwen3vl_merger`, `projector:step3vl`, `projector:ultravox`, `projector:voxtral`, `projector:yasa2`, `projector:youtuvl` | metadata schema; tensor closure; component graph parity |
 | `intentionally-rejected` | `policy-rejections` | `architecture:bailingmoe2`, `architecture:clip`, `architecture:dots3note`, `architecture:exaone-moe`, `architecture:exaone4`, `architecture:glm4`, `architecture:glm4moe`, `architecture:gptj` | policy change plus independent correctness proof |
 | `intentionally-rejected` | `policy-rejections` | `projector:pockettts_gen`, `projector:qwen3tts_gen` | sidecar role must become a valid projector contract |
@@ -199,7 +198,7 @@ Reason codes are concise user-facing categories; detailed architecture audits re
 | `gemma3n` | — | none (fails before config extraction) | exact-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — Gemma3n GGUF is the text member of a vision-and-audio package whose gemma3nv and gemma3na clip companions carry distinct encoders and projectors. |
 | `gemma4` | — | model=`gemma4_text`; tensor=`llama`+`gemma4_extras`; mmproj=`gemma4` | not claimed | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
 | `gemma4-assistant` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — Gemma4 Assistant is a standalone target-coupled model with pre/post projections, masked embeddings, scalar layer scales, its own KV cache, and a live target-model context. |
-| `glm-dsa` | `glm_dsa` | none (no tensor mapping route) | audited-direct-loader-conditional-union | config=supported; tensor_map=deferred; graph=supported; runtime=deferred; quantized_import=supported | TENSOR_MAP_DEFERRED — 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. |
+| `glm-dsa` | `glm_dsa` | model=`glm_moe_dsa`; tensor=`glm_dsa` | audited-direct-loader-conditional-union | config=supported; tensor_map=supported; graph=supported; runtime=deferred; quantized_import=supported | RUNTIME_EVIDENCE_PENDING — Config extraction, exact tensor-name closure, and a full synthetic GGUF graph build are covered, but no representative real-weight GGUF has yet passed ORT parity or generation validation. |
 | `glm4` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — GLM4 serializes complete fused-FFN trailing blocks and NextN tensors, but the pinned loader skips appended blocks; GLM-OCR converter transforms also permute Q/K for M-RoPE. |
 | `glm4moe` | — | none (fails before config extraction) | audited-direct-loader-conditional-union | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — GLM4-MoE serializes biased attention and periodic dense/routed expert trailing blocks with mandatory router bias, but the pinned loader skips them. |
 | `gpt-oss` | — | none (fails before config extraction) | not claimed | config=deferred; tensor_map=deferred; graph=deferred; runtime=deferred; quantized_import=supported | CONFIG_DEFERRED — The pinned GPT-OSS converter splits interleaved gate/up expert rows and repacks checkpoint block+scale tensors into expert-major MXFP4 values. |

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

@@ -2079,13 +2079,11 @@
         model_type="glm_moe_dsa",
         aliases=frozenset({"glm_dsa"}),
         config_key_map="glm_dsa",
-        tensor_map=Support.DEFERRED,
-        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
-        ),
+        config_postprocessor="glm_dsa",
+        tensor_map_recipe=("glm_dsa",),
+        tensor_processor="glm_dsa",
+        rope_interleave=True,
+        reason=_RUNTIME_VALIDATION_PENDING,
     ),
     GGUFArchitectureSpec(
         gguf_arch="apertus",

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 = 99
+_EXPECTED_SUPPORTED_COUNT = 100
 _PROMOTED_CONVENTIONAL_DECODERS = frozenset(
     {
         "bitnet",
@@ -190,6 +190,7 @@
         "gemma3",
         "gemma4",
         "granite",
+        "glm-dsa",
         "granitemoe",
         "hunyuan-dense",
         "jamba",