From 690b87d88151dbd775bb6eff5581663688785928 Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Sat, 21 Nov 2026 03:10:38 -0800 Subject: [PATCH] model : inkling2_loop share the loop slots after the scale pass (#32704) The loop slots copied the physical layers at the end of load_arch_tensors, before the generic pass attaches the optional .scale tensors, so the later passes through the core ran NVFP4 experts without their per-expert scales. Add an optional post_arch_tensors hook that runs after the scale pass and share the slots there. --- src/llama-model.cpp | 1 + src/llama-model.h | 3 +++ src/models/inkling2-loop.cpp | 2 ++ src/models/models.h | 1 + 4 files changed, 7 insertions(+) diff --git a/src/llama-model.cpp b/src/llama-model.cpp index 4c77f6f68..81e28b1c5 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -2019,6 +2019,7 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) { output_in_s = create_tensor(tn(LLM_TENSOR_OUTPUT, "input_scale"), {1}, TENSOR_NOT_REQUIRED); } } + post_arch_tensors(ml); } ml.done_getting_tensors(); diff --git a/src/llama-model.h b/src/llama-model.h index e86e5e54f..6473980a8 100644 --- a/src/llama-model.h +++ b/src/llama-model.h @@ -1070,6 +1070,9 @@ struct llama_model { virtual void load_arch_tensors(llama_model_loader & ml) = 0; virtual std::unique_ptr build_arch_graph(const llm_graph_params & params) const = 0; + // optional, runs after the scale pass + virtual void post_arch_tensors(llama_model_loader &) {} + protected: llama_model_params params; diff --git a/src/models/inkling2-loop.cpp b/src/models/inkling2-loop.cpp index 0de66baee..8366a1191 100644 --- a/src/models/inkling2-loop.cpp +++ b/src/models/inkling2-loop.cpp @@ -60,7 +60,9 @@ void llama_model_inkling2_loop::load_arch_tensors(llama_model_loader & ml) { l.ctrl_forget[c] = create_tensor(tn(forget[c], "weight", il), {rank, n_embd}, 0); } } +} +void llama_model_inkling2_loop::post_arch_tensors(llama_model_loader &) { // share the physical weights, each cache slot keeps its own KV and conv state for (int il = n_phys; il < (int) hparams.n_layer(); ++il) { layers[il] = layers[layer_phys(il)]; diff --git a/src/models/models.h b/src/models/models.h index 636afeae9..8d2a58827 100644 --- a/src/models/models.h +++ b/src/models/models.h @@ -2547,6 +2547,7 @@ struct llama_model_inkling2_loop : public llama_model_inkling2 { llama_model_inkling2_loop(const struct llama_model_params & params) : llama_model_inkling2(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; + void post_arch_tensors(llama_model_loader & ml) override; int n_layer_phys() const override { return n_phys; } int layer_phys(int il) const;