toy 0.8.0

data/lib/toy/llm/engine/llama_seq_engine_cuda.rb

@@ -0,0 +1,1238 @@
+# lib/toy/llm/engine/llama_seq_engine_cuda.rb — CUDA mirror of lib/toy/llm/engine/llama_seq_engine.rb.
+#
+# AUTO-GENERATED by prep/gen_cuda_mirror.rb. Do not edit by hand;
+# edit the CPU source and re-run the generator. The CPU file's
+# header explains the architecture; this mirror keeps the same
+# contract on the GPU backend via TinyNNCuda.
+# llama-family architecture (RMSNorm + SwiGLU + RoPE + GQA).
+#
+# Sibling of lib/toy_smollm2_ffi_kv.rb (KV-cache decode for single
+# positions). This class builds ONE forward graph that processes T
+# token positions in a single compute — the prerequisite for real SFT
+# (F1.2 step 6c/6d) where masked CE across all T positions is the loss.
+#
+# Step 1 (this file): forward only, no LoRA, no training. Parity gate
+# is T=1 vs SmolLM2KVFFICache#decode_step at pos=0; full-T parity smoke
+# (T>1) follows in step 2.
+#
+# Spinel hygiene — see docs/design/m3-seq-forward-2026-05-21.md, risk #6:
+# the M1 full-forward integration was blocked by ivar-type unification
+# across cache classes (t_seq / d_model / d_ff fields colliding). To
+# sidestep it cleanly, every dimension/parameter ivar here uses a
+# `@seq_*` prefix. Verbose but type-isolated.
+
+require_relative "../../models/transformer"
+require_relative "../../../toy"
+require_relative "../../models/toy_smollm2"
+require_relative "../../ffi/tinynn_cuda"
+require_relative "../primitives/rms_norm_cuda"
+require_relative "../primitives/rope_cuda"
+require_relative "../primitives/swiglu_cuda"
+require_relative "../primitives/gqa_cuda"
+require_relative "../blocks/transformer_block_cuda"
+require_relative "../archs/llama_arch_cuda"
+
+module Toy; module LLM; module Engine
+class LlamaSeqEngineCuda
+  attr_accessor :sess,
+                # P2.5 — the five arch-level persistent handles
+                # (t_seq_token_embed, t_seq_final_norm_gamma, t_seq_output,
+                # t_seq_w_proj) and the blocks array now LIVE on @seq_arch
+                # (Toy::LLM::Archs::LlamaArch). Cache delegators below
+                # preserve the public accessor surface so the realize
+                # paths, external PCA-init (fcache.t_seq_w_proj=), and the
+                # examples (fcache.t_seq_*) keep working by name.
+                :seq_arch,
+                :seq_has_untied_output, :seq_has_qkv_bias,
+                :seq_t, :seq_b, :seq_d_model, :seq_d_ff, :seq_n_heads, :seq_n_kv,
+                :seq_d_head, :seq_group_size, :seq_n_layers, :seq_vocab_size,
+                :seq_rope_base, :seq_rope_scaling, :t_seq_rope_freq_factors,
+                :seq_rms_eps, :seq_realized, :seq_weight_dtype,
+                :t_seq_token_ids, :t_seq_positions, :t_seq_attn_mask,
+                :t_seq_x_embed, :t_seq_x_final, :t_seq_logits,
+                :seq_gguf_handle_keepalive,
+                # M3 step 3 LoRA flags. Both must be set BEFORE
+                # realize_for_mmap. enable_lora_q!(r) allocates the
+                # adapter A/B pair per Q head; enable_lora_q_adamw!
+                # additionally allocates persistent Adam m/v alongside
+                # them (mirrors F1.2 step 6b).
+                :seq_lora_q_enabled, :seq_lora_q_rank,
+                :seq_lora_q_adamw_enabled,
+                # F3 full fine-tune. When set BEFORE
+                # realize_for_full_finetune, every per-block weight
+                # tensor is allocated writable F32 in ctx_w with paired
+                # Adam m/v and marked set_param; build_training_step
+                # then emits opt_step on each.
+                :seq_full_finetune_enabled,
+                # Per-cache (not per-block) parallel arrays for the
+                # global trainable weights — token_embed, final-norm
+                # gamma, optional untied output projection. Populated
+                # in realize_for_full_finetune; empty otherwise.
+                :ft_globals_weights, :ft_globals_m, :ft_globals_v,
+                # Opt-in: include the embedding / final-norm /
+                # untied-output tensors in full FT. Default off — the
+                # embed tensor on Qwen-class models is huge (V=152K ×
+                # d=1536 = 233M params F32) and brings the full
+                # forward+backward+opt-step memory closer to the GPU
+                # limit. Works correctly on CUDA after vendor-patches
+                # 0006 chunked the get_rows_back kernel launch
+                # (previously hit CUDA's gridDim.y = 65535 cap).
+                :ft_train_embeddings_enabled
+
+  # P2.5 — delegators forwarding the arch-owned handle accessors to
+  # @seq_arch (Toy::LLM::Archs::LlamaArch). These preserve the cache's
+  # former public attr_accessor surface (the realize paths assign via
+  # self.t_seq_token_embed=, external PCA-init writes fcache.t_seq_w_proj=,
+  # examples read fcache.t_seq_*). Single source of truth: the arch.
+  def t_seq_token_embed;        @seq_arch.t_seq_token_embed;        end
+  def t_seq_token_embed=(v);    @seq_arch.t_seq_token_embed = v;    end
+  def t_seq_final_norm_gamma;     @seq_arch.t_seq_final_norm_gamma;     end
+  def t_seq_final_norm_gamma=(v); @seq_arch.t_seq_final_norm_gamma = v; end
+  def t_seq_output;             @seq_arch.t_seq_output;             end
+  def t_seq_output=(v);         @seq_arch.t_seq_output = v;         end
+  def t_seq_w_proj;             @seq_arch.t_seq_w_proj;             end
+  def t_seq_w_proj=(v);         @seq_arch.t_seq_w_proj = v;         end
+  # E2.3 — projection-lens donor width (0 disables the lens). Plain
+  # ivar (NOT in the attr_accessor list); the GGUF-fold writer reads it
+  # to know the donor->d_model contraction dimension.
+  def seq_donor_d_in;           @seq_donor_d_in;                    end
+  def seq_blocks_ffi;           @seq_arch.seq_blocks_ffi;           end
+  def seq_blocks_ffi=(v);       @seq_arch.seq_blocks_ffi = v;       end
+
+  def initialize
+    # P2.5 — the arch owns the arch-level persistent handles + the
+    # blocks array (seeded with one block in the arch ctor, matching the
+    # former cache seed). Constructed first so the delegators are live.
+    @seq_arch       = Toy::LLM::Archs::LlamaArch.new
+    @seq_realized   = false
+    @seq_t          = 0
+    @seq_b          = 1
+    # GH#9 — mixed-precision compute. 0 = F32 (current behaviour;
+    # bit-identical to pre-GH#9). 1 = F16, 30 = BF16. When != 0,
+    # weight matmuls inside build_seq_block / build_seq_qhead route
+    # through mp_matmul which casts the F32 master to the chosen
+    # dtype inline in the forward graph. F32 master is kept (required
+    # by opt_step_adamw); the cast result lives in transient scratch.
+    @seq_weight_dtype = 0
+    @seq_d_model    = 0
+    @seq_d_ff       = 0
+    @seq_n_heads    = 0
+    @seq_n_kv       = 0
+    @seq_d_head     = 0
+    @seq_group_size = 0
+    @seq_n_layers   = 0
+    @seq_vocab_size = 0
+    @seq_rope_base            = 10000.0
+    @seq_rope_scaling         = Toy::RopeScaling.none
+    # Seed a concrete Cfg from the same defaults so the ivar always
+    # holds a real Toy::LLM::Primitives::RoPE::Cfg (never nil/RbVal).
+    # Rebuilt per realize path once the true dims are known.
+    @seq_rope_cfg             = Toy::LLM::Primitives::RoPE::Cfg.new(
+                                  @seq_d_head, @seq_rope_base,
+                                  @seq_rope_scaling.freq_scale,
+                                  @seq_rope_scaling.ext_factor,
+                                  @seq_rope_scaling.attn_factor,
+                                  @seq_rope_scaling.beta_fast,
+                                  @seq_rope_scaling.beta_slow)
+    @t_seq_rope_freq_factors  = TinyNNCuda.tnn_null_ptr
+    @seq_rms_eps    = 1.0e-5
+    @sess                  = TinyNNCuda.tnn_null_ptr
+    # P2.5 — token_embed / final_norm_gamma / output / w_proj and the
+    # blocks array are seeded on @seq_arch (see arch ctor); the cache
+    # reaches them via the delegators above.
+    @seq_has_untied_output = false
+    @seq_has_qkv_bias      = false
+    @seq_gguf_handle_keepalive = TinyNNCuda.tnn_null_ptr
+    @t_seq_token_ids = TinyNNCuda.tnn_null_ptr
+    @t_seq_positions = TinyNNCuda.tnn_null_ptr
+    # GH#7 — batched-training block-causal attention mask. Allocated
+    # only when @seq_b > 1 (realize_for_random_init with t_batch > 1);
+    # otherwise stays NULL and build_seq_qhead falls back to the
+    # diag_mask_inf + softmax path (bit-identical to today at B=1).
+    @t_seq_attn_mask = TinyNNCuda.tnn_null_ptr
+    @t_seq_x_embed   = TinyNNCuda.tnn_null_ptr
+    @t_seq_x_final   = TinyNNCuda.tnn_null_ptr
+    @t_seq_logits    = TinyNNCuda.tnn_null_ptr
+    @seq_lora_q_enabled       = false
+    @seq_lora_q_rank          = 0
+    @seq_lora_q_adamw_enabled = false
+    @seq_full_finetune_enabled = false
+    @ft_globals_weights = [TinyNNCuda.tnn_null_ptr]; @ft_globals_weights.pop
+    @ft_globals_m       = [TinyNNCuda.tnn_null_ptr]; @ft_globals_m.pop
+    @ft_globals_v       = [TinyNNCuda.tnn_null_ptr]; @ft_globals_v.pop
+    @ft_train_embeddings_enabled = false
+    # E2.3 (towards GH#14) — projection-lens path. donor_d_in is read
+    # from cfg in realize_for_random_init; t_seq_w_proj is the
+    # trainable [donor_d_in, d_model] linear inserted after the embed
+    # get_rows when donor_d_in > 0.
+    @seq_donor_d_in   = 0
+    # P2.5 — t_seq_w_proj is seeded on @seq_arch (arch ctor); cache
+    # reaches it via the t_seq_w_proj delegator.
+  end
+
+  # F3 — additionally train the embedding / final-norm gamma / untied
+  # output. Opt-in: the embed tensor on Qwen-class vocab is large and
+  # makes the memory budget noticeably tighter, but the math itself
+  # works correctly post vendor-patches/0006 (chunked get_rows_back).
+  def enable_full_finetune_embeddings!
+    @ft_train_embeddings_enabled = true
+  end
+
+  # F3 — turn on full fine-tune. Every per-block weight tensor will
+  # be allocated as writable F32 in ctx_w (instead of mmap'd from the
+  # GGUF), paired with persistent Adam m/v, and marked trainable.
+  # Mutually exclusive with enable_lora_q!. Call BEFORE
+  # realize_for_full_finetune.
+  def enable_full_finetune!
+    @seq_full_finetune_enabled = true
+  end
+
+  # M3 step 3 — turn on LoRA on the Q projection. Adapter A is (r, d_model),
+  # B is (d_head, r). Standard init: A small Gaussian, B zero → adapter is
+  # a no-op at step 0. Call BEFORE realize_for_mmap. Mirrors
+  # SmolLM2KVFFICache#enable_lora_q!.
+  def enable_lora_q!(r)
+    @seq_lora_q_enabled = true
+    @seq_lora_q_rank    = r
+  end
+
+  # M3 step 3 — allocate persistent AdamW moments next to each LoRA pair
+  # (parallel to F1.2 step 6b on SmolLM2KVFFICache). Required to keep
+  # optimizer state alive across reset_for_rebuild / multi-step training.
+  def enable_lora_q_adamw!
+    @seq_lora_q_adamw_enabled = true
+  end
+
+  # P2.6 — shared config-prologue helper. Writes the @seq_* shape/RoPE
+  # ivars that every realize_for_* path needs before allocating tensors.
+  # Pure ivar writes reading only cfg.*; no FFI, no graph state. Each
+  # realize path keeps its own `@seq_t = t_seq` (and any path-local
+  # extras) at the call site and then calls this. Byte-identical to the
+  # block that previously lived inline in all four realize_for_* methods.
+  def apply_seq_cfg!(cfg)
+    @seq_d_model    = cfg.d_model
+    @seq_d_ff       = cfg.d_ff
+    @seq_n_heads    = cfg.n_heads
+    @seq_n_kv       = cfg.n_kv
+    @seq_d_head     = cfg.head_dim
+    @seq_group_size = cfg.n_heads / cfg.n_kv
+    @seq_n_layers   = cfg.n_layers
+    @seq_vocab_size = cfg.vocab
+    @seq_rope_base    = cfg.rope_base
+    @seq_rope_scaling = cfg.rope_scaling
+    @seq_rope_cfg     = Toy::LLM::Primitives::RoPE::Cfg.new(
+                          @seq_d_head, @seq_rope_base,
+                          @seq_rope_scaling.freq_scale,
+                          @seq_rope_scaling.ext_factor,
+                          @seq_rope_scaling.attn_factor,
+                          @seq_rope_scaling.beta_fast,
+                          @seq_rope_scaling.beta_slow)
+    @seq_rms_eps    = cfg.rms_eps
+  end
+
+  # F4 alternative realize for CUDA + Q8 base. Allocates every weight
+  # tensor in the standard ggml ctx_w (NOT the BYO mmap region), then
+  # verbatim-copies the GGUF bytes in. Buys correctness on CUDA at the
+  # cost of holding the weights twice transiently (mmap + ctx_w during
+  # load; ctx_w only after). Required because the BYO-pointer cuda
+  # buffer's quantized padding zeroing (cudaMemset past tensor data)
+  # would otherwise crash on Q8 tensors with `ne0 % 512 != 0`.
+  #
+  # Use this realize when (a) the GGUF is Q8 AND (b) the backend is
+  # CUDA. CPU + Q8 stays on realize_for_mmap (no padding issue).
+  def realize_for_q8_copy(gguf_handle, cfg, t_seq, untied, qkv_bias)
+    @seq_t          = t_seq
+    apply_seq_cfg!(cfg)
+
+    @seq_gguf_handle_keepalive = gguf_handle
+    @sess                  = TinyNNCuda.tnn_session_new(1)
+
+    # llama3 / LongRoPE: allocate the freq_factors tensor in ctx_w
+    # before finalize_weights. Values uploaded post-finalize.
+    if @seq_rope_scaling.kind == :llama3
+      @t_seq_rope_freq_factors = TinyNNCuda.tnn_rope_freq_factors_alloc(@sess, cfg.head_dim)
+    else
+      @t_seq_rope_freq_factors = TinyNNCuda.tnn_null_ptr
+    end
+    @seq_has_untied_output = untied
+    @seq_has_qkv_bias      = qkv_bias
+
+    # Read source tensor types so we can allocate ctx_w tensors of the
+    # MATCHING type (verbatim copy requires source/target types match).
+    eidx = TinyNNCuda.tnn_gguf_find_index(gguf_handle, "token_embd.weight")
+    etyp = TinyNNCuda.tnn_gguf_tensor_type(gguf_handle, eidx)
+    self.t_seq_token_embed = TinyNNCuda.tnn_input_2d_persistent_typed(@sess,
+                           @seq_vocab_size, @seq_d_model, etyp)
+    self.t_seq_final_norm_gamma = TinyNNCuda.tnn_input_1d_f32_persistent(@sess, @seq_d_model)
+    if untied
+      oidx = TinyNNCuda.tnn_gguf_find_index(gguf_handle, "output.weight")
+      otyp = TinyNNCuda.tnn_gguf_tensor_type(gguf_handle, oidx)
+      self.t_seq_output = TinyNNCuda.tnn_input_2d_persistent_typed(@sess,
+                        @seq_vocab_size, @seq_d_model, otyp)
+    end
+
+    # P2.6 Step 2 — seeding loop moved onto the arch (LlamaArch#seed_blocks!).
+    @seq_arch.seed_blocks!(@seq_n_layers)
+
+    # P2.7 pass-3 — the per-block ALLOC-typed loop body moved onto
+    # TransformerBlock#alloc_q8_typed_from_gguf! (verbatim; called ONLY
+    # from here). Mirrors load_from_gguf_mmap!'s arg-passing exactly: every
+    # dim/flag arrives as an arg, NO ivar reads off the block. The q8 path
+    # never names LoRA tensors, so the moved body is :str-free (#16-clean)
+    # — no lora_name_q! back-calls (unlike load_from_gguf_mmap!).
+    li = 0
+    while li < @seq_n_layers
+      blk = self.seq_blocks_ffi[li]
+      blk.alloc_q8_typed_from_gguf!(@sess, gguf_handle, li,
+                                    @seq_n_heads, @seq_n_kv, @seq_d_head, @seq_d_model,
+                                    @seq_d_ff, @seq_vocab_size, @seq_lora_q_enabled,
+                                    @seq_lora_q_rank, @seq_lora_q_adamw_enabled, qkv_bias)
+      li = li + 1
+    end
+
+    if @seq_lora_q_enabled
+      li2 = 0
+      while li2 < @seq_n_layers
+        blk2 = self.seq_blocks_ffi[li2]
+        hq_p = 0
+        while hq_p < @seq_n_heads
+          TinyNNCuda.tnn_set_param(blk2.t_seq_w_lora_a_q[hq_p])
+          TinyNNCuda.tnn_set_param(blk2.t_seq_w_lora_b_q[hq_p])
+          hq_p = hq_p + 1
+        end
+        li2 = li2 + 1
+      end
+    end
+
+    finalize_weights_and_upload_constants!
+
+    # Load all weight bytes from the GGUF into the now-allocated
+    # backend buffers. Verbatim copy keeps Q8 as Q8.
+    TinyNNCuda.tnn_gguf_copy_verbatim_to_persistent(gguf_handle, eidx, @sess, self.t_seq_token_embed)
+    fnidx = TinyNNCuda.tnn_gguf_find_index(gguf_handle, "output_norm.weight")
+    TinyNNCuda.tnn_gguf_copy_1d_to_persistent(gguf_handle, fnidx, @sess, self.t_seq_final_norm_gamma)
+    if untied
+      oidx2 = TinyNNCuda.tnn_gguf_find_index(gguf_handle, "output.weight")
+      TinyNNCuda.tnn_gguf_copy_verbatim_to_persistent(gguf_handle, oidx2, @sess, self.t_seq_output)
+    end
+
+    # P2.7 pass-3 Step 2 — the per-block VERBATIM-COPY loop body moved onto
+    # TransformerBlock#copy_q8_bytes_from_gguf! (verbatim; called ONLY from
+    # here). The copy phase fills the backend buffers allocated by the
+    # alloc_q8_typed_from_gguf! pass; the block reads its OWN t_seq_* handles
+    # and writes nothing on itself. NO ivar reads off the cache — every dim
+    # (n_heads, n_kv, d_head) and the qkv_bias flag arrive as ARGS. All the
+    # moved primitives are tnn_gguf_copy_* / tnn_gguf_find_index — the same
+    # :str-at-runtime pattern alloc_q8_typed_from_gguf! already uses, never
+    # block class-load scope (#16). The GLOBALS verbatim-copy above (token
+    # embed / final norm / untied output) STAYS on the cache — those touch
+    # cache-level t_seq_* handles, not the block.
+    li_l = 0
+    while li_l < @seq_n_layers
+      blk = self.seq_blocks_ffi[li_l]
+      blk.copy_q8_bytes_from_gguf!(@sess, gguf_handle, li_l,
+                                   @seq_n_heads, @seq_n_kv, @seq_d_head, qkv_bias)
+      li_l = li_l + 1
+    end
+
+    if @seq_lora_q_adamw_enabled
+      za = Mat.new(@seq_lora_q_rank, @seq_d_model)
+      zb = Mat.new(@seq_d_head,      @seq_lora_q_rank)
+      i = 0
+      while i < @seq_lora_q_rank * @seq_d_model; za.flat[i] = 0.0; i = i + 1; end
+      j = 0
+      while j < @seq_d_head * @seq_lora_q_rank; zb.flat[j] = 0.0; j = j + 1; end
+      li_z = 0
+      while li_z < @seq_n_layers
+        blk_z = self.seq_blocks_ffi[li_z]
+        hqz = 0
+        while hqz < @seq_n_heads
+          TinyNNCuda.upload_row_major(@sess, blk_z.t_seq_w_lora_a_q_m[hqz], za)
+          TinyNNCuda.upload_row_major(@sess, blk_z.t_seq_w_lora_a_q_v[hqz], za)
+          TinyNNCuda.upload_row_major(@sess, blk_z.t_seq_w_lora_b_q_m[hqz], zb)
+          TinyNNCuda.upload_row_major(@sess, blk_z.t_seq_w_lora_b_q_v[hqz], zb)
+          hqz = hqz + 1
+        end
+        li_z = li_z + 1
+      end
+    end
+
+    build_and_realize!
+  end
+
+  # Allocate persistent weights mmap'd from `gguf_handle` (caller is
+  # responsible for keeping the handle alive — we keepalive it via
+  # @seq_gguf_handle_keepalive), compute inputs, and the full forward
+  # graph for T = `t_seq` positions. Fixed T; rebuild for a different T.
+  #
+  # Weight layout matches SmolLM2KVFFICache#realize_for_mmap exactly
+  # (same byte offsets + per-head split), so a sharded GGUF can be
+  # loaded by either class.
+  def realize_for_mmap(gguf_handle, cfg, t_seq, untied, qkv_bias)
+    @seq_t          = t_seq
+    apply_seq_cfg!(cfg)
+
+    @seq_gguf_handle_keepalive = gguf_handle
+    @sess                  = TinyNNCuda.tnn_session_new(1)
+
+    # llama3 / LongRoPE: allocate the freq_factors tensor in ctx_w
+    # before finalize_weights. Values uploaded post-finalize.
+    if @seq_rope_scaling.kind == :llama3
+      @t_seq_rope_freq_factors = TinyNNCuda.tnn_rope_freq_factors_alloc(@sess, cfg.head_dim)
+    else
+      @t_seq_rope_freq_factors = TinyNNCuda.tnn_null_ptr
+    end
+    @seq_has_untied_output = untied
+    @seq_has_qkv_bias      = qkv_bias
+
+    map_base = TinyNNCuda.tnn_gguf_mmap_base(gguf_handle)
+    map_size = TinyNNCuda.tnn_gguf_mmap_size(gguf_handle)
+    TinyNNCuda.tnn_session_attach_weight_mmap(@sess, map_base, map_size)
+
+    # Embeddings + final norm + optional untied LM head.
+    # P2.6 pass-2 Step 1 — the three arch-owned global mmap allocs moved
+    # onto LlamaArch#load_globals_from_gguf_mmap! (verbatim; called ONLY
+    # from here). Mirrors the seed_blocks! / alloc_trainable_f32_weights!
+    # extraction precedents.
+    @seq_arch.load_globals_from_gguf_mmap!(@sess, gguf_handle,
+                                           @seq_vocab_size, @seq_d_model, untied)
+
+    # P2.6 Step 2 — seeding loop moved onto the arch (LlamaArch#seed_blocks!).
+    @seq_arch.seed_blocks!(@seq_n_layers)
+
+    # P2.7 — the per-block alloc-from-mmap-offsets loop body moved onto
+    # TransformerBlock#load_from_gguf_mmap! (verbatim; called ONLY from
+    # here). Mirrors the alloc_trainable_f32_weights! / seed_blocks! /
+    # load_globals_from_gguf_mmap! extraction precedents. head_nbytes and
+    # the LoRA :str tnn_tensor_set_name naming stay on THIS cache and are
+    # back-called through the passed `self` ref (lora_name_q! /
+    # lora_name_q_adam! issue the :str FFI at this runtime scope, never in
+    # block class-load scope — landmine #16).
+    li = 0
+    while li < @seq_n_layers
+      blk = self.seq_blocks_ffi[li]
+      blk.load_from_gguf_mmap!(@sess, self, gguf_handle, li,
+                               @seq_n_heads, @seq_n_kv, @seq_d_head, @seq_d_model,
+                               @seq_d_ff, @seq_lora_q_enabled, @seq_lora_q_rank,
+                               @seq_lora_q_adamw_enabled, qkv_bias)
+      li = li + 1
+    end
+
+    # Mark LoRA tensors as trainable BEFORE finalize. set_param flips
+    # the PARAM flag so build_backward walks them when emitting grad nodes.
+    if @seq_lora_q_enabled
+      li2 = 0
+      while li2 < @seq_n_layers
+        blk2 = self.seq_blocks_ffi[li2]
+        hq_p = 0
+        while hq_p < @seq_n_heads
+          TinyNNCuda.tnn_set_param(blk2.t_seq_w_lora_a_q[hq_p])
+          TinyNNCuda.tnn_set_param(blk2.t_seq_w_lora_b_q[hq_p])
+          hq_p = hq_p + 1
+        end
+        li2 = li2 + 1
+      end
+    end
+
+    finalize_weights_and_upload_constants!
+
+    # Zero-init persistent AdamW moments. Same contract as F1.2 step 6b
+    # on SmolLM2KVFFICache — m and v start at 0 per the AdamW update rule.
+    if @seq_lora_q_adamw_enabled
+      za = Mat.new(@seq_lora_q_rank, @seq_d_model)
+      zb = Mat.new(@seq_d_head,      @seq_lora_q_rank)
+      i = 0
+      while i < @seq_lora_q_rank * @seq_d_model; za.flat[i] = 0.0; i = i + 1; end
+      j = 0
+      while j < @seq_d_head * @seq_lora_q_rank; zb.flat[j] = 0.0; j = j + 1; end
+      li_z = 0
+      while li_z < @seq_n_layers
+        blk_z = self.seq_blocks_ffi[li_z]
+        hqz = 0
+        while hqz < @seq_n_heads
+          TinyNNCuda.upload_row_major(@sess, blk_z.t_seq_w_lora_a_q_m[hqz], za)
+          TinyNNCuda.upload_row_major(@sess, blk_z.t_seq_w_lora_a_q_v[hqz], za)
+          TinyNNCuda.upload_row_major(@sess, blk_z.t_seq_w_lora_b_q_m[hqz], zb)
+          TinyNNCuda.upload_row_major(@sess, blk_z.t_seq_w_lora_b_q_v[hqz], zb)
+          hqz = hqz + 1
+        end
+        li_z = li_z + 1
+      end
+    end
+
+    build_and_realize!
+  end
+
+  # F3 — full fine-tune realize path. Parallel to realize_for_mmap
+  # but every per-block weight is allocated writable F32 in ctx_w
+  # (no mmap), set_param-marked, paired with Adam m/v, and loaded
+  # from the GGUF post-finalize via the dequantize-friendly
+  # tnn_gguf_copy_* primitives. The embedding tensor + final_norm
+  # gamma stay mmap'd (read-only) — the MVP doesn't train them.
+  def realize_for_full_finetune(gguf_handle, cfg, t_seq, untied, qkv_bias)
+    @seq_t          = t_seq
+    apply_seq_cfg!(cfg)
+
+    @seq_gguf_handle_keepalive = gguf_handle
+    @sess                  = TinyNNCuda.tnn_session_new(1)
+
+    # llama3 / LongRoPE: allocate the freq_factors tensor in ctx_w
+    # before finalize_weights. Values uploaded post-finalize.
+    if @seq_rope_scaling.kind == :llama3
+      @t_seq_rope_freq_factors = TinyNNCuda.tnn_rope_freq_factors_alloc(@sess, cfg.head_dim)
+    else
+      @t_seq_rope_freq_factors = TinyNNCuda.tnn_null_ptr
+    end
+    @seq_has_untied_output = untied
+    @seq_has_qkv_bias      = qkv_bias
+
+    # Token embed + final-norm gamma + (untied) output: trainable
+    # only when opt-in (ft_train_embeddings_enabled). Otherwise
+    # they stay mmap'd / read-only (still need a mmap attach for
+    # this branch).
+    if @ft_train_embeddings_enabled
+      self.t_seq_token_embed = TinyNNCuda.tnn_input_2d_f32_persistent(@sess,
+                             @seq_vocab_size, @seq_d_model)
+      ft_add_global_2d(self.t_seq_token_embed, @seq_vocab_size, @seq_d_model)
+      ft_name_last_global("token_embd.weight")
+
+      self.t_seq_final_norm_gamma = TinyNNCuda.tnn_input_1d_f32_persistent(@sess, @seq_d_model)
+      ft_add_global_1d(self.t_seq_final_norm_gamma)
+      ft_name_last_global("output_norm.weight")
+
+      if untied
+        self.t_seq_output = TinyNNCuda.tnn_input_2d_f32_persistent(@sess,
+                          @seq_vocab_size, @seq_d_model)
+        ft_add_global_2d(self.t_seq_output, @seq_vocab_size, @seq_d_model)
+        ft_name_last_global("output.weight")
+      end
+    else
+      map_base = TinyNNCuda.tnn_gguf_mmap_base(gguf_handle)
+      map_size = TinyNNCuda.tnn_gguf_mmap_size(gguf_handle)
+      TinyNNCuda.tnn_session_attach_weight_mmap(@sess, map_base, map_size)
+
+      eidx = TinyNNCuda.tnn_gguf_find_index(gguf_handle, "token_embd.weight")
+      eoff = TinyNNCuda.tnn_gguf_tensor_file_offset(gguf_handle, eidx)
+      etyp = TinyNNCuda.tnn_gguf_tensor_type(gguf_handle, eidx)
+      self.t_seq_token_embed = TinyNNCuda.tnn_input_2d_persistent_mmap(@sess,
+                             @seq_vocab_size, @seq_d_model, etyp, eoff)
+
+      fnidx = TinyNNCuda.tnn_gguf_find_index(gguf_handle, "output_norm.weight")
+      fnoff = TinyNNCuda.tnn_gguf_tensor_file_offset(gguf_handle, fnidx)
+      self.t_seq_final_norm_gamma = TinyNNCuda.tnn_input_1d_persistent_mmap(@sess,
+                                  @seq_d_model, 0, fnoff)
+
+      if untied
+        oidx = TinyNNCuda.tnn_gguf_find_index(gguf_handle, "output.weight")
+        ooff = TinyNNCuda.tnn_gguf_tensor_file_offset(gguf_handle, oidx)
+        otyp = TinyNNCuda.tnn_gguf_tensor_type(gguf_handle, oidx)
+        self.t_seq_output = TinyNNCuda.tnn_input_2d_persistent_mmap(@sess,
+                          @seq_vocab_size, @seq_d_model, otyp, ooff)
+      end
+    end
+
+    # P2.6 Step 2 — seeding loop moved onto the arch (LlamaArch#seed_blocks!).
+    @seq_arch.seed_blocks!(@seq_n_layers)
+
+    # P2-finish — per-block FT alloc lifted onto the block (verbatim:
+    # TransformerBlock#alloc_full_finetune_f32_weights!), mirroring how
+    # realize_for_random_init drives alloc_trainable_f32_weights!. The block
+    # owns its self.t_seq_* handles + the per-block set_param loop; the cache
+    # passes @sess + dims + qkv_bias and the ft_add_*/ft_name_last recorders
+    # are back-called. Gated byte-exact by prep/full_finetune_gate.rb.
+    li = 0
+    while li < @seq_n_layers
+      blk = self.seq_blocks_ffi[li]
+      prefix = "blk." + li.to_s + "."
+      blk.alloc_full_finetune_f32_weights!(@sess, self, prefix,
+                                           @seq_d_model, @seq_d_ff, @seq_d_head,
+                                           @seq_n_heads, @seq_n_kv, qkv_bias)
+      li = li + 1
+    end
+
+    # Globals are trainable too only when embeddings are opt-in.
+    if @ft_train_embeddings_enabled
+      gi = 0
+      while gi < @ft_globals_weights.length
+        TinyNNCuda.tnn_set_param(@ft_globals_weights[gi])
+        gi = gi + 1
+      end
+    end
+
+    finalize_weights_and_upload_constants!
+
+    # Post-finalize: load every writable weight from the GGUF.
+    if @ft_train_embeddings_enabled
+      ft_load_globals(gguf_handle, untied)
+    end
+    ft_load_from_gguf(gguf_handle, qkv_bias)
+    ft_zero_init_adam(qkv_bias)
+    if @ft_train_embeddings_enabled
+      ft_zero_init_adam_globals
+    end
+
+    build_and_realize!
+  end
+
+  # P2-α: from-scratch training entry. Allocates the same persistent
+  # tensor layout as realize_for_full_finetune (embeddings + per-block
+  # weights, all trainable F32 in ctx_w), then random-initialises
+  # every weight via Ruby-side Gaussian upload — no GGUF needed.
+  #
+  # Force-enables `@ft_train_embeddings_enabled` so the existing
+  # full-FT machinery allocates persistent F32 embeddings instead
+  # of the mmap branch. Caller doesn't need to call
+  # enable_full_finetune_embeddings! first.
+  #
+  # Currently Llama-arch only (RMSNorm + GQA + RoPE + SwiGLU). Other
+  # architectures (GPT-2 LN, MHA + biases) need a separate trainer
+  # cache class; deferred until we actually need GPT-2 from-scratch.
+  def realize_for_random_init(cfg, t_seq, t_batch, weight_dtype, untied, qkv_bias, seed, init_scale)
+    @ft_train_embeddings_enabled = true   # forces persistent-F32 alloc of embeddings
+    @seq_full_finetune_enabled   = true   # build_training_step gates on this
+
+    @seq_t          = t_seq
+    # GH#7 — micro-batching. B=1 keeps the codepath bit-identical to
+    # the pre-GH#7 single-sequence training. B>1 lays out tokens as a
+    # flat [T*B] vector with a block-causal attention mask uploaded
+    # post-finalize and applied via soft_max_ext.
+    @seq_b          = t_batch
+    # GH#9 — mixed-precision compute. 0 = F32 (bit-identical to
+    # pre-GH#9). 1 = F16, 30 = BF16. See mp_matmul + ivar comment in
+    # initialize for the master-copy details.
+    @seq_weight_dtype = weight_dtype
+    apply_seq_cfg!(cfg)
+
+    @sess                  = TinyNNCuda.tnn_session_new(1)
+    # GH#17 — per-head decomposition makes node count scale as
+    # O(n_layers × n_heads). The default 65536 cap overflows on
+    # 24L × 16-head Qwen-shape at backward-expand. Empirically a
+    # 24L × 16-head model needs ~450k nodes for forward + backward +
+    # AdamW, so we budget ~1000 nodes per (layer × head) cell + floor.
+    cap = cfg.n_layers * cfg.n_heads * 1000 + 65536
+    TinyNNCuda.tnn_session_set_graph_capacity(@sess, cap)
+    @seq_has_untied_output = untied
+    @seq_has_qkv_bias      = qkv_bias
+    @seq_donor_d_in        = cfg.donor_d_in   # E2.3 — 0 disables projection lens
+
+    if @seq_rope_scaling.kind == :llama3
+      @t_seq_rope_freq_factors = TinyNNCuda.tnn_rope_freq_factors_alloc(@sess, cfg.head_dim)
+    else
+      @t_seq_rope_freq_factors = TinyNNCuda.tnn_null_ptr
+    end
+
+    # Globals — trainable persistent F32 (+ E2.3 projection-lens branch when
+    # @seq_donor_d_in > 0). P2-finish: the alloc lifted onto the arch
+    # (LlamaArch#alloc_globals_trainable_f32!), which already owns these handles
+    # — verbatim, same order, byte-identical. @ft_globals_* recorders + the
+    # frozen-embed namer are back-called through `self`.
+    @seq_arch.alloc_globals_trainable_f32!(@sess, self, @seq_vocab_size,
+                                           @seq_d_model, @seq_donor_d_in, untied)
+
+    # Per-block weights — identical structure to realize_for_full_finetune.
+    # P2.6 Step 2 — the block-array seeding loop now lives on the arch
+    # (LlamaArch#seed_blocks!), which already owns @seq_blocks_ffi.
+    @seq_arch.seed_blocks!(@seq_n_layers)
+
+    # P2.6 Step 4 — the per-block F32 ALLOC loop body now lives on the
+    # block (TransformerBlock#alloc_trainable_f32_weights!), which already
+    # OWNS these self.t_seq_* handles at forward time. The block takes
+    # @sess + the seq dims + the name prefix as ARGS (no ivar reads on the
+    # block) and calls the cache's ft_add_1d / ft_add_2d / ft_name_last
+    # recorders BACK through the passed `self` reference — those stay on
+    # the cache (they read @sess and issue tnn_tensor_set_name :str at
+    # runtime; never migrate into block class-load scope). w_o keeps its
+    # random_init shape ne=[d_model, n_heads*d_head] inside the block
+    # method (not unified with full_finetune's [d_model,d_model]).
+    li = 0
+    while li < @seq_n_layers
+      blk = self.seq_blocks_ffi[li]
+      prefix = "blk." + li.to_s + "."
+      blk.alloc_trainable_f32_weights!(@sess, self, prefix,
+                                       @seq_d_model, @seq_d_ff, @seq_d_head,
+                                       @seq_n_heads, @seq_n_kv)
+      li = li + 1
+    end
+
+    # Mark globals as params too (gated on @ft_train_embeddings_enabled).
+    gi = 0
+    while gi < @ft_globals_weights.length
+      TinyNNCuda.tnn_set_param(@ft_globals_weights[gi])
+      gi = gi + 1
+    end
+
+    finalize_weights_and_upload_constants!
+
+    # Random-init every weight + zero biases + ones gammas.
+    upload_random_init!(seed, init_scale, qkv_bias, untied)
+    ft_zero_init_adam(qkv_bias)
+    ft_zero_init_adam_globals
+
+    build_and_realize!
+  end
+
+  # GH#7 — build + upload the block-causal attention mask for B>1.
+  # Layout: scores from matmul(K[d_head, T*B], Q[d_head, T*B]) have
+  # ne=[T*B, T*B] where ne0 indexes keys and ne1 indexes queries
+  # (ggml column-major: flat[ne0_idx + ne1_idx * T*B]). For query
+  # position i1 = b_q*T + p_q and key position i0 = b_k*T + p_k:
+  #   mask = 0.0  iff b_k == b_q AND p_k <= p_q   (intra-batch causal)
+  #   mask = NEG  otherwise                      (cross-batch + future)
+  # NEG = -1.0e30 so exp(NEG) == 0.0 in f32 (avoids Float::INFINITY,
+  # which would also work but is one less Spinel codegen variable).
+  def upload_block_causal_mask!
+    tb = @seq_t * @seq_b
+    neg = -1.0e30
+    mask_arr = [0.0]; mask_arr.pop
+    i1 = 0
+    while i1 < tb
+      b_q = i1 / @seq_t
+      p_q = i1 % @seq_t
+      i0 = 0
+      while i0 < tb
+        b_k = i0 / @seq_t
+        p_k = i0 % @seq_t
+        if b_k == b_q && p_k <= p_q
+          mask_arr.push(0.0)
+        else
+          mask_arr.push(neg)
+        end
+        i0 = i0 + 1
+      end
+      i1 = i1 + 1
+    end
+    TinyNNCuda.tnn_upload_from_float_array(@sess, @t_seq_attn_mask, mask_arr, mask_arr.length)
+  end
+
+  # Fill every persistent weight tensor with N(0, std) values.
+  # Norm gammas → 1.0, biases (if present) → 0.0, matmul weights →
+  # N(0, init_scale/sqrt(fan_in)). Token embedding uses GPT-2-style
+  # N(0, 0.02). All values computed in Ruby, uploaded in bulk via
+  # tnn_upload_from_float_array.
+  def upload_random_init!(seed, init_scale, qkv_bias, untied)
+    state = [seed]
+
+    # Token embed: width depends on projection lens.
+    # When donor_d_in > 0, embed is [vocab, donor_d_in] (caller may
+    # overwrite with real donor values after realize); the trainable
+    # projection W_proj [donor_d_in, d_model] also gets a Gaussian init.
+    embed_cols = @seq_donor_d_in > 0 ? @seq_donor_d_in : @seq_d_model
+    upload_gaussian(self.t_seq_token_embed, @seq_vocab_size * embed_cols, 0.02, state)
+    if @seq_donor_d_in > 0
+      upload_gaussian(self.t_seq_w_proj, @seq_donor_d_in * @seq_d_model,
+                       1.0 / Math.sqrt(@seq_donor_d_in.to_f), state)
+    end
+    upload_constant(self.t_seq_final_norm_gamma, @seq_d_model, 1.0)
+    if untied
+      upload_gaussian(self.t_seq_output, @seq_vocab_size * @seq_d_model, 0.02, state)
+    end
+
+    inv_sqrt_d   = init_scale / Math.sqrt(@seq_d_model.to_f)
+    inv_sqrt_dff = init_scale / Math.sqrt(@seq_d_ff.to_f)
+
+    li = 0
+    while li < @seq_n_layers
+      blk = self.seq_blocks_ffi[li]
+      upload_constant(blk.t_seq_rn1_gamma, @seq_d_model, 1.0)
+      upload_constant(blk.t_seq_rn2_gamma, @seq_d_model, 1.0)
+
+      hq = 0
+      while hq < @seq_n_heads
+        upload_gaussian(blk.t_seq_w_q[hq], @seq_d_head * @seq_d_model, inv_sqrt_d, state)
+        hq = hq + 1
+      end
+      hkv = 0
+      while hkv < @seq_n_kv
+        upload_gaussian(blk.t_seq_w_k[hkv], @seq_d_head * @seq_d_model, inv_sqrt_d, state)
+        upload_gaussian(blk.t_seq_w_v[hkv], @seq_d_head * @seq_d_model, inv_sqrt_d, state)
+        hkv = hkv + 1
+      end
+
+      upload_gaussian(blk.t_seq_w_o,    @seq_d_model * @seq_n_heads * @seq_d_head, inv_sqrt_d, state)
+      upload_gaussian(blk.t_seq_w_gate, @seq_d_ff    * @seq_d_model, inv_sqrt_d, state)
+      upload_gaussian(blk.t_seq_w_up,   @seq_d_ff    * @seq_d_model, inv_sqrt_d, state)
+      upload_gaussian(blk.t_seq_w_down, @seq_d_model * @seq_d_ff,    inv_sqrt_dff, state)
+      li = li + 1
+    end
+  end
+
+  # Box-Muller from a xorshift64-driven uniform stream. state is a
+  # one-element Array<Integer> so the mutable PRNG state survives
+  # across calls without using class variables. Always emits exactly
+  # `n` Gaussian-distributed F32 values via tnn_upload_from_float_array.
+  def upload_gaussian(tensor, n, std, state)
+    buf = [0.0]; buf.pop
+    pair = 0
+    saved = 0.0
+    i = 0
+    while i < n
+      if pair == 0
+        u1 = xorshift_uniform!(state)
+        u2 = xorshift_uniform!(state)
+        if u1 < 1.0e-300; u1 = 1.0e-300; end
+        r = Math.sqrt(-2.0 * Math.log(u1))
+        theta = 2.0 * Math::PI * u2
+        z0 = r * Math.cos(theta) * std
+        z1 = r * Math.sin(theta) * std
+        buf.push(z0)
+        saved = z1
+        pair = 1
+      else
+        buf.push(saved)
+        pair = 0
+      end
+      i = i + 1
+    end
+    TinyNNCuda.tnn_upload_from_float_array(@sess, tensor, buf, n)
+  end
+
+  def upload_constant(tensor, n, v)
+    buf = [0.0]; buf.pop
+    i = 0
+    while i < n
+      buf.push(v)
+      i = i + 1
+    end
+    TinyNNCuda.tnn_upload_from_float_array(@sess, tensor, buf, n)
+  end
+
+  # xorshift64 → uniform in (0, 1). Mutates state[0].
+  def xorshift_uniform!(state)
+    x = state[0]
+    x = x ^ (x << 13)
+    x = x & 0xFFFFFFFFFFFFFFFF
+    x = x ^ (x >> 7)
+    x = x ^ (x << 17)
+    x = x & 0xFFFFFFFFFFFFFFFF
+    state[0] = x
+    (x.to_f / 18446744073709551616.0) + 1.0e-300
+  end
+
+  # Append (weight, m, v) to the block's parallel arrays. Allocates
+  # Adam m and v of the same shape as `weight` as a side effect.
+  def ft_add_2d(blk, weight, rows, cols)
+    blk.ft_weights.push(weight)
+    blk.ft_m.push(TinyNNCuda.tnn_input_2d_f32_persistent(@sess, rows, cols))
+    blk.ft_v.push(TinyNNCuda.tnn_input_2d_f32_persistent(@sess, rows, cols))
+  end
+
+  def ft_add_1d(blk, weight)
+    n = TinyNNCuda.tnn_tensor_nelements(weight)
+    blk.ft_weights.push(weight)
+    blk.ft_m.push(TinyNNCuda.tnn_input_1d_f32_persistent(@sess, n))
+    blk.ft_v.push(TinyNNCuda.tnn_input_1d_f32_persistent(@sess, n))
+  end
+
+  # Name the most-recently-pushed (weight, m, v) triple in a block.
+  # Used right after ft_add_2d / ft_add_1d so drift/grad event consumers
+  # see llama.cpp-convention names like "blk.0.attn_norm.weight" instead
+  # of ggml's auto-generated "node_N". toy#semantic-tensor-names.
+  def ft_name_last(blk, name)
+    last = blk.ft_weights.length - 1
+    TinyNNCuda.tnn_tensor_set_name(blk.ft_weights[last], name)
+    TinyNNCuda.tnn_tensor_set_name(blk.ft_m[last],       name + ".m")
+    TinyNNCuda.tnn_tensor_set_name(blk.ft_v[last],       name + ".v")
+  end
+
+  def ft_name_last_global(name)
+    last = @ft_globals_weights.length - 1
+    TinyNNCuda.tnn_tensor_set_name(@ft_globals_weights[last], name)
+    TinyNNCuda.tnn_tensor_set_name(@ft_globals_m[last],       name + ".m")
+    TinyNNCuda.tnn_tensor_set_name(@ft_globals_v[last],       name + ".v")
+  end
+
+  # Name a single FROZEN global (e.g. the projection-lens donor embed, which is
+  # NOT pushed to @ft_globals so ft_name_last_global cannot reach it). Kept on
+  # the engine so this tnn_tensor_set_name(:str) FFI stays on the cache realize
+  # runtime path — same discipline as ft_name_last / lora_name_q!. Back-called
+  # by LlamaArch#alloc_globals_trainable_f32!.
+  def name_global!(t, name)
+    TinyNNCuda.tnn_tensor_set_name(t, name)
+  end
+
+  # P2.7 — LoRA-Q tensor naming callbacks for the extracted block-side
+  # mmap loader (TransformerBlock#load_from_gguf_mmap!). The :str
+  # tnn_tensor_set_name FFI calls MUST stay on the cache realize RUNTIME
+  # path — never migrate into block class-load scope (step_bind / :str
+  # landmine #16). The block assembles the runtime name string and hands
+  # it here, exactly as it hands ft_name_last its assembled name. Verbatim
+  # lift of the former realize_for_mmap loop lines 567-570 / 597-604.
+  def lora_name_q!(t_a, t_b, head_prefix)
+    TinyNNCuda.tnn_tensor_set_name(t_a, head_prefix + ".lora_a.weight")
+    TinyNNCuda.tnn_tensor_set_name(t_b, head_prefix + ".lora_b.weight")
+  end
+
+  def lora_name_q_adam!(t_a_m, t_a_v, t_b_m, t_b_v, head_prefix)
+    TinyNNCuda.tnn_tensor_set_name(t_a_m, head_prefix + ".lora_a.m")
+    TinyNNCuda.tnn_tensor_set_name(t_a_v, head_prefix + ".lora_a.v")
+    TinyNNCuda.tnn_tensor_set_name(t_b_m, head_prefix + ".lora_b.m")
+    TinyNNCuda.tnn_tensor_set_name(t_b_v, head_prefix + ".lora_b.v")
+  end
+
+  # Same shape as ft_add_2d / ft_add_1d but writes to the cache-level
+  # globals arrays (token_embed, final-norm, untied output).
+  def ft_add_global_2d(weight, rows, cols)
+    @ft_globals_weights.push(weight)
+    @ft_globals_m.push(TinyNNCuda.tnn_input_2d_f32_persistent(@sess, rows, cols))
+    @ft_globals_v.push(TinyNNCuda.tnn_input_2d_f32_persistent(@sess, rows, cols))
+  end
+
+  def ft_add_global_1d(weight)
+    n = TinyNNCuda.tnn_tensor_nelements(weight)
+    @ft_globals_weights.push(weight)
+    @ft_globals_m.push(TinyNNCuda.tnn_input_1d_f32_persistent(@sess, n))
+    @ft_globals_v.push(TinyNNCuda.tnn_input_1d_f32_persistent(@sess, n))
+  end
+
+  # Load token_embed + final-norm + (untied) output from the GGUF
+  # into their now-allocated backend buffers.
+  def ft_load_globals(gguf, untied)
+    eidx = TinyNNCuda.tnn_gguf_find_index(gguf, "token_embd.weight")
+    TinyNNCuda.tnn_gguf_copy_to_persistent(gguf, eidx, @sess, self.t_seq_token_embed)
+    fnidx = TinyNNCuda.tnn_gguf_find_index(gguf, "output_norm.weight")
+    TinyNNCuda.tnn_gguf_copy_1d_to_persistent(gguf, fnidx, @sess, self.t_seq_final_norm_gamma)
+    if untied
+      oidx = TinyNNCuda.tnn_gguf_find_index(gguf, "output.weight")
+      TinyNNCuda.tnn_gguf_copy_to_persistent(gguf, oidx, @sess, self.t_seq_output)
+    end
+  end
+
+  def ft_zero_init_adam_globals
+    gi = 0
+    while gi < @ft_globals_weights.length
+      TinyNNCuda.tnn_zero_tensor(@sess, @ft_globals_m[gi])
+      TinyNNCuda.tnn_zero_tensor(@sess, @ft_globals_v[gi])
+      gi = gi + 1
+    end
+  end
+
+  # Pull bytes from the GGUF into each writable weight. Uses the
+  # existing C-side dequantize-and-copy primitives so a Q8 source
+  # transparently becomes F32 in the target tensor.
+  def ft_load_from_gguf(gguf, qkv_bias)
+    li = 0
+    while li < @seq_n_layers
+      blk = self.seq_blocks_ffi[li]
+      prefix = "blk." + li.to_s
+
+      rn1_idx = TinyNNCuda.tnn_gguf_find_index(gguf, prefix + ".attn_norm.weight")
+      rn2_idx = TinyNNCuda.tnn_gguf_find_index(gguf, prefix + ".ffn_norm.weight")
+      TinyNNCuda.tnn_gguf_copy_1d_to_persistent(gguf, rn1_idx, @sess, blk.t_seq_rn1_gamma)
+      TinyNNCuda.tnn_gguf_copy_1d_to_persistent(gguf, rn2_idx, @sess, blk.t_seq_rn2_gamma)
+
+      q_idx = TinyNNCuda.tnn_gguf_find_index(gguf, prefix + ".attn_q.weight")
+      hq = 0
+      while hq < @seq_n_heads
+        TinyNNCuda.tnn_gguf_copy_head_slice_to_persistent_native(gguf, q_idx, @sess,
+          blk.t_seq_w_q[hq], hq, @seq_n_heads, @seq_d_model, @seq_d_head)
+        hq = hq + 1
+      end
+
+      k_idx = TinyNNCuda.tnn_gguf_find_index(gguf, prefix + ".attn_k.weight")
+      v_idx = TinyNNCuda.tnn_gguf_find_index(gguf, prefix + ".attn_v.weight")
+      hkv = 0
+      while hkv < @seq_n_kv
+        TinyNNCuda.tnn_gguf_copy_head_slice_to_persistent_native(gguf, k_idx, @sess,
+          blk.t_seq_w_k[hkv], hkv, @seq_n_kv, @seq_d_model, @seq_d_head)
+        TinyNNCuda.tnn_gguf_copy_head_slice_to_persistent_native(gguf, v_idx, @sess,
+          blk.t_seq_w_v[hkv], hkv, @seq_n_kv, @seq_d_model, @seq_d_head)
+        hkv = hkv + 1
+      end
+
+      if qkv_bias
+        # qbias / kbias / vbias are 1-D head-sliced. We don't have a
+        # dedicated head-slice loader for them; fall through and use
+        # tnn_gguf_copy_head_bias_slice_to_persistent.
+        qb_idx = TinyNNCuda.tnn_gguf_find_index(gguf, prefix + ".attn_q.bias")
+        kb_idx = TinyNNCuda.tnn_gguf_find_index(gguf, prefix + ".attn_k.bias")
+        vb_idx = TinyNNCuda.tnn_gguf_find_index(gguf, prefix + ".attn_v.bias")
+        hbq = 0
+        while hbq < @seq_n_heads
+          TinyNNCuda.tnn_gguf_copy_head_bias_slice_to_persistent(gguf, qb_idx, @sess,
+            blk.t_seq_b_q[hbq], hbq, @seq_d_head)
+          hbq = hbq + 1
+        end
+        hbkv = 0
+        while hbkv < @seq_n_kv
+          TinyNNCuda.tnn_gguf_copy_head_bias_slice_to_persistent(gguf, kb_idx, @sess,
+            blk.t_seq_b_k[hbkv], hbkv, @seq_d_head)
+          TinyNNCuda.tnn_gguf_copy_head_bias_slice_to_persistent(gguf, vb_idx, @sess,
+            blk.t_seq_b_v[hbkv], hbkv, @seq_d_head)
+          hbkv = hbkv + 1
+        end
+      end
+
+      o_idx    = TinyNNCuda.tnn_gguf_find_index(gguf, prefix + ".attn_output.weight")
+      gate_idx = TinyNNCuda.tnn_gguf_find_index(gguf, prefix + ".ffn_gate.weight")
+      up_idx   = TinyNNCuda.tnn_gguf_find_index(gguf, prefix + ".ffn_up.weight")
+      down_idx = TinyNNCuda.tnn_gguf_find_index(gguf, prefix + ".ffn_down.weight")
+      TinyNNCuda.tnn_gguf_copy_to_persistent(gguf, o_idx,    @sess, blk.t_seq_w_o)
+      TinyNNCuda.tnn_gguf_copy_to_persistent(gguf, gate_idx, @sess, blk.t_seq_w_gate)
+      TinyNNCuda.tnn_gguf_copy_to_persistent(gguf, up_idx,   @sess, blk.t_seq_w_up)
+      TinyNNCuda.tnn_gguf_copy_to_persistent(gguf, down_idx, @sess, blk.t_seq_w_down)
+
+      li = li + 1
+    end
+  end
+
+  # Zero-init the Adam moments. m and v both start at 0 per the
+  # AdamW step-0 contract. Uses the backend-side memset primitive
+  # (tnn_zero_tensor) so a 1 GB Adam state doesn't materialize a
+  # Mat-of-zeros in Ruby first.
+  def ft_zero_init_adam(qkv_bias)
+    li = 0
+    while li < @seq_n_layers
+      blk = self.seq_blocks_ffi[li]
+      i = 0
+      while i < blk.ft_weights.length
+        TinyNNCuda.tnn_zero_tensor(@sess, blk.ft_m[i])
+        TinyNNCuda.tnn_zero_tensor(@sess, blk.ft_v[i])
+        i = i + 1
+      end
+      li = li + 1
+    end
+  end
+
+  # P2.6 — finalize the backend weight buffers and upload the
+  # per-model constants that depend on the buffers existing. This is
+  # the identical head-of-tail shared by all four realize_for_* paths:
+  #   1. allocate the B>1 block-causal mask in ctx_w (NULL at B=1),
+  #   2. tnn_finalize_weights,
+  #   3. upload the llama3 RoPE freq_factors (no-op unless :llama3),
+  #   4. upload the B>1 block-causal mask values.
+  # Stays a CACHE method: the finalize FFI sequencing is session-scoped.
+  # Gate-covered end-to-end by smoke_projection_lens (B=1, non-llama3):
+  # the two inner branches are dead under the gate but relocate verbatim.
+  def finalize_weights_and_upload_constants!
+    # GH#7 — block-causal attention mask for B>1. At B=1 the mask stays
+    # NULL and build_seq_qhead uses diag_mask_inf + softmax. Allocated
+    # in ctx_w as f32 persistent so it survives reset_for_rebuild.
+    if @seq_b > 1
+      tb_alloc = @seq_t * @seq_b
+      @t_seq_attn_mask = TinyNNCuda.tnn_input_2d_f32_persistent(@sess, tb_alloc, tb_alloc)
+    end
+
+    TinyNNCuda.tnn_finalize_weights(@sess)
+
+    # Upload llama3-style RoPE freq_factors once the backend buffer
+    # exists. Per-model constant; never re-uploaded.
+    if @seq_rope_scaling.kind == :llama3
+      ff = Toy::RopeScaling.compute_llama3_freq_factors(
+        @seq_d_head, @seq_rope_base,
+        @seq_rope_scaling.orig_max_pos, @seq_rope_scaling.factor,
+        @seq_rope_scaling.low_freq_factor, @seq_rope_scaling.high_freq_factor)
+      TinyNNCuda.tnn_upload_from_float_array(@sess, @t_seq_rope_freq_factors, ff, ff.length)
+    end
+
+    if @seq_b > 1
+      upload_block_causal_mask!
+    end
+  end
+
+  # P2.6 — the identical tail-of-tail shared by all four realize_for_*
+  # paths: build the forward graph in the current ctx, realize it, and
+  # flip @seq_realized. Stays a CACHE method (build_forward_in_current_ctx
+  # is the cache->arch wrapper; tnn_realize is session-scoped).
+  # Gate-covered by smoke_projection_lens via realize_for_random_init.
+  def build_and_realize!
+    build_forward_in_current_ctx
+    TinyNNCuda.tnn_realize(@sess, @t_seq_logits)
+    @seq_realized = true
+  end
+
+  # Build the forward graph in the CURRENT compute context. Used both
+  # from realize_for_mmap (first realize) and after tnn_reset_for_rebuild
+  # (e.g. when switching from inference to training, which needs the
+  # forward + loss + backward + opt_step all in one rebuilt ctx).
+  # Stores the per-graph tensor handles back on `self`.
+  # P2.5 — thin wrapper around Toy::LLM::Archs::LlamaArch#build_forward.
+  # Allocates the per-graph INPUT handles (token_ids, positions) — which
+  # stay CACHE-owned graph I/O, read by forward() and the uploaders —
+  # then hands the realize-set rope_cfg / donor_d_in onto the arch and
+  # calls the lifted orchestration. The three per-graph OUTPUT handles
+  # come back in a LlamaArchForwardOut and are spread onto the cache's
+  # own ivars so every downstream reader (@t_seq_logits accessor,
+  # build_training_step CE-loss consumer, examples/06 fcache.t_seq_logits)
+  # is untouched.
+  def build_forward_in_current_ctx
+    # GH#7 — at B=1, @seq_t * @seq_b == @seq_t (legacy behaviour).
+    # At B>1, the layout is flat [T*B]: per-batch positions cycle
+    # 0..T-1 (the caller-built positions array is responsible for
+    # that ordering); RoPE applies per-batch positional encoding
+    # because rope_ext reads positions[k] for each ne[2] slot.
+    tb = @seq_t * @seq_b
+    @t_seq_token_ids = TinyNNCuda.tnn_input_1d_i32(@sess, tb)
+    @t_seq_positions = TinyNNCuda.tnn_input_1d_i32_ctx(@sess, tb)
+
+    # The arch reads seq_rope_cfg / seq_donor_d_in off itself; the cache
+    # rebuilds rope_cfg and sets donor_d_in in each realize prologue, so
+    # mirror the realize-set values onto the arch right before the call.
+    @seq_arch.seq_rope_cfg   = @seq_rope_cfg
+    @seq_arch.seq_donor_d_in = @seq_donor_d_in
+
+    out = @seq_arch.build_forward(
+      @sess, @t_seq_token_ids, @t_seq_positions, @t_seq_rope_freq_factors,
+      @t_seq_attn_mask, @seq_rms_eps, @seq_d_head, @seq_n_kv, @seq_n_heads,
+      @seq_group_size, @seq_has_qkv_bias, @seq_weight_dtype,
+      @seq_lora_q_enabled, @seq_t, @seq_b, @seq_n_layers,
+      @seq_has_untied_output)
+
+    @t_seq_x_embed = out.t_seq_x_embed
+    @t_seq_x_final = out.t_seq_x_final
+    @t_seq_logits  = out.t_seq_logits
+  end
+
+  # M3 step 3 — rebuild the session graph as forward + CE loss + backward
+  # + AdamW opt_step over every LoRA pair. After this, callers upload
+  # token IDs + positions + labels (one-hot vocab×T) + hp vector and
+  # call tnn_compute_backward to get one training step over the whole
+  # T-position sequence.
+  #
+  # Returns the (loss_tensor, labels_tensor, hp_tensor) triple.
+  def build_training_step
+    if !@seq_full_finetune_enabled && (!@seq_lora_q_enabled || !@seq_lora_q_adamw_enabled)
+      puts "build_training_step: requires enable_lora_q! AND enable_lora_q_adamw!  (or enable_full_finetune!)"
+      return nil
+    end
+    TinyNNCuda.tnn_reset_for_rebuild(@sess)
+    build_forward_in_current_ctx
+
+    # Label tensor: same shape as logits, ggml ne=[vocab, T*B]. Our
+    # wrapper takes (rows, cols) and emits ggml(cols, rows), so pass
+    # (T*B, vocab) here to get ne=[vocab, T*B]. One-hot per ne1-column
+    # (i.e. per (batch, position) slot). At B=1, identical to legacy.
+    t_labels = TinyNNCuda.tnn_input_2d_f32(@sess, @seq_t * @seq_b, @seq_vocab_size)
+    # Hyper-params vector for AdamW: alpha, beta1, beta2, eps, wd, beta1h, beta2h.
+    t_hp = TinyNNCuda.tnn_input_1d_f32(@sess, 7)
+
+    # CE loss over all T columns. ggml_cross_entropy_loss returns the
+    # mean over columns — masking is a follow-up (would zero specific
+    # columns in labels before this op).
+    t_loss = TinyNNCuda.tnn_cross_entropy_loss(@sess, @t_seq_logits, t_labels)
+    TinyNNCuda.tnn_set_output(t_loss)
+    TinyNNCuda.tnn_set_loss(t_loss)
+
+    TinyNNCuda.tnn_build_forward_only(@sess, t_loss)
+    TinyNNCuda.tnn_build_backward(@sess)
+
+    if @seq_full_finetune_enabled
+      # F3 — emit opt_step_adamw for every recorded (weight, m, v)
+      # triple. The arrays are populated in realize_for_full_finetune.
+      li = 0
+      while li < @seq_n_layers
+        blk = self.seq_blocks_ffi[li]
+        wi = 0
+        while wi < blk.ft_weights.length
+          tw = blk.ft_weights[wi]
+          tg = TinyNNCuda.tnn_tensor_grad(@sess, tw)
+          to = TinyNNCuda.tnn_opt_step_adamw(@sess, tw, tg,
+                                          blk.ft_m[wi], blk.ft_v[wi], t_hp)
+          TinyNNCuda.tnn_extend_backward_graph(@sess, to)
+          wi = wi + 1
+        end
+        li = li + 1
+      end
+      # Globals (token_embed, final-norm, optional untied output).
+      gi = 0
+      while gi < @ft_globals_weights.length
+        tw = @ft_globals_weights[gi]
+        tg = TinyNNCuda.tnn_tensor_grad(@sess, tw)
+        to = TinyNNCuda.tnn_opt_step_adamw(@sess, tw, tg,
+                                        @ft_globals_m[gi], @ft_globals_v[gi], t_hp)
+        TinyNNCuda.tnn_extend_backward_graph(@sess, to)
+        gi = gi + 1
+      end
+    else
+      # LoRA-only training (M3 step 3). One opt_step_adamw per LoRA-A
+      # and per LoRA-B tensor; thread each through extend_backward_graph
+      # so sched sees the writes.
+      li = 0
+      while li < @seq_n_layers
+        blk = self.seq_blocks_ffi[li]
+        hq = 0
+        while hq < @seq_n_heads
+          t_a       = blk.t_seq_w_lora_a_q[hq]
+          t_b       = blk.t_seq_w_lora_b_q[hq]
+          t_grad_a  = TinyNNCuda.tnn_tensor_grad(@sess, t_a)
+          t_grad_b  = TinyNNCuda.tnn_tensor_grad(@sess, t_b)
+          t_opt_a   = TinyNNCuda.tnn_opt_step_adamw(@sess, t_a, t_grad_a,
+                                                  blk.t_seq_w_lora_a_q_m[hq],
+                                                  blk.t_seq_w_lora_a_q_v[hq], t_hp)
+          t_opt_b   = TinyNNCuda.tnn_opt_step_adamw(@sess, t_b, t_grad_b,
+                                                  blk.t_seq_w_lora_b_q_m[hq],
+                                                  blk.t_seq_w_lora_b_q_v[hq], t_hp)
+          TinyNNCuda.tnn_extend_backward_graph(@sess, t_opt_a)
+          TinyNNCuda.tnn_extend_backward_graph(@sess, t_opt_b)
+          hq = hq + 1
+        end
+        li = li + 1
+      end
+    end
+
+    # Pin every node in graph_b before sched-alloc — workaround for the
+    # ggml-cpu sched-aliasing bug on long backward chains (documented in
+    # project_cpu_cuda_lora_train_divergence_2026_05_21). Memory cost
+    # grows roughly with node count; fine for SmolLM2-135M at T<=64.
+    TinyNNCuda.tnn_pin_all_graph_b_nodes(@sess)
+    TinyNNCuda.tnn_realize_backward(@sess)
+    [t_loss, t_labels, t_hp]
+  end
+
+  # GGUF type → bytes-per-row stride for per-head slicing. Mirrors the
+  # SmolLM2KVFFICache helper of the same name. F32=0, Q8_0=8.
+  def head_nbytes(ggml_type, d_head, d_model)
+    if ggml_type == 0
+      d_head * d_model * 4
+    elsif ggml_type == 8
+      d_head * (d_model / 32) * 34
+    else
+      0
+    end
+  end
+
+  # Run one forward pass. `ids` and `positions` are length-T Int arrays.
+  # Returns the t_seq_logits handle; caller downloads via download_row_major
+  # against (vocab, T) shape.
+  def forward(ids, positions)
+    TinyNNCuda.upload_int_array(@sess, @t_seq_token_ids, ids)
+    TinyNNCuda.upload_int_array(@sess, @t_seq_positions, positions)
+    TinyNNCuda.tnn_compute(@sess)
+    @t_seq_logits
+  end
+
+  # Seed LoRA-A with a small Gaussian and LoRA-B with zero — the
+  # standard init makes the adapter a no-op at step 0 (forward output
+  # equals the base model). Mirror of SmolLM2KVFFICache#upload_lora_q_init!.
+  def upload_lora_q_init!(seed, init_scale)
+    if !@seq_lora_q_enabled; return; end
+    s = seed
+    m_a = Mat.new(@seq_lora_q_rank, @seq_d_model)
+    m_b = Mat.new(@seq_d_head, @seq_lora_q_rank)
+    i_b = 0
+    while i_b < @seq_d_head * @seq_lora_q_rank
+      m_b.flat[i_b] = 0.0
+      i_b = i_b + 1
+    end
+    li = 0
+    while li < @seq_n_layers
+      blk = self.seq_blocks_ffi[li]
+      hq = 0
+      while hq < @seq_n_heads
+        ii = 0
+        while ii < @seq_lora_q_rank * @seq_d_model
+          s = (s * 1103515245 + 12345) & 0x7FFFFFFF
+          u1 = (s.to_f + 1.0) / 2147483648.0
+          s = (s * 1103515245 + 12345) & 0x7FFFFFFF
+          u2 = (s.to_f + 1.0) / 2147483648.0
+          m_a.flat[ii] = init_scale * Math.sqrt(-2.0 * Math.log(u1)) * Math.cos(2.0 * Math::PI * u2)
+          ii = ii + 1
+        end
+        TinyNNCuda.upload_row_major(@sess, blk.t_seq_w_lora_a_q[hq], m_a)
+        TinyNNCuda.upload_row_major(@sess, blk.t_seq_w_lora_b_q[hq], m_b)
+        hq = hq + 1
+      end
+      li = li + 1
+    end
+  end
+end
+end; end; end