Fix glm5next MoE hybrid placement vs config

[marcelormendes]

Problem

GLM-5.3-Flash dual-GPU initialization was failing in nine stages:

Stages 1-8 ✅ FIXED

Previous fixes for placement, CPU reads, contexts, layer detection, get_rows workaround, etc.

Stage 9: Missing attn_wo binding + graph hash overflow ✅ FIXED

Issue 1: All 46 layers skipped attention

layer 0 ... (mla: all nil, kda: f_a/f_b/g_a/g_b SET, wo=(nil))
layer 3 ... (mla: q_a/q_b/wk_b/wv_b SET, wo=(nil), kda: all nil)
...all 46 layers have wo=(nil)

Cause: GGUF doesn't have blk.N.attn_out.weight. Loader only tried one name, so attn_wo was null on every layer. This caused has_kda and has_mla checks (which require attn_wo) to fail, forcing all layers to skip attention and feed garbage into MoE.

Issue 2: Graph hash overflow

ggml.c:7307: GGML_ASSERT(node_hash_pos != GGML_HASHSET_FULL) failed

Cause: Default ggml_new_graph() hash size (2048) is too small for 46-layer × 288-expert MoE model.

Solution

Fixes 1-8 (commits 9c70323 through 3b564cd)

See previous sections.

Fix 9: Bind attn_wo with fallbacks + increase graph size (f93f12b)

1. Multiple attn_wo naming patterns:

Try common GGUF tensor name variations:

layer.attn_wo = ggml_get_tensor(ctx, "blk.N.attn_out.weight");
if (!layer.attn_wo) 
    layer.attn_wo = ggml_get_tensor(ctx, "blk.N.attn_output.weight");
if (!layer.attn_wo) 
    layer.attn_wo = ggml_get_tensor(ctx, "blk.N.attn_wo.weight");

Applied to both KDA and MLA loader branches so all attention layers can find their output projection.

2. Increase graph hash capacity:

// Before: default size 2048
ggml_cgraph * gf = ggml_new_graph(ctx);

// After: 32768 nodes for large MoE
ggml_cgraph * gf = ggml_new_graph_custom(ctx, 32768, false);

Applied to both prefill and decode graph creation. Size matches laguna's allocation for large MoE models.

Expected result:

• Layers 0-2: detect as KDA (have kda_f_a/f_b/g_a/g_b + attn_wo)
• Layers 3,7,11,...: detect as MLA (have attn_q_a/q_b/wk_b/wv_b + attn_wo)
• Other layers: detect as KDA
• All 46 layers process attention correctly
• Graph build completes without hash overflow

Testing

Status on soulf:

• ✅ All 9 stages should pass
• 🔄 Awaiting smoke test: Proper attention processing and token generation

Model: GLM-5.3-Flash skip-list v3 (46 layers, 288 experts).

Commits

1. 9c70323 - Fix placement configuration
2. ebdf7f2 - CPU expert tensor reads
3. 4f54d75 - Buffer getter compatibility
4. 45b12f1 - Layer type detection
5. f6a4671 - Require attn_wo check
6. 5b0ff88 - Separate graph/const contexts
7. 3df9659 - Add const_ctx to MLA
8. 3b564cd - Workaround get_rows + diagnostics
9. f93f12b - Fix attn_wo binding + graph hash size