rtx/src/c--/scheduler.cmm

-- =====================================================================
-- SOV SCHEDULER -- C-- Continuous Batching State Machine
-- States: IDLE(0) PREFILL(1) GENERATE(2) SWAP(3) CHECKPOINT(4) RESUME(5)
-- Janet array: 32 Word32 slots at known address for runtime config
-- WORM: every 64 generated tokens → worm_checkpoint() → Blake3+Ed25519 receipt
-- =====================================================================

section "data" {
    -- Janet array: 32 Word32 slots
    -- [0]  = pending_request_count
    -- [1]  = batch_size
    -- [2]  = tokens_generated (mod 64 counter)
    -- [3]  = current_seq_id
    -- [4]  = kv_blocks_used
    -- [5]  = power_state (0=active,1=suspend,2=resume,3=low_battery)
    -- [6]  = speculative_draft_len
    -- [7]  = bft_quorum_height
    -- [8..15] = worm_receipt_blake3 (8 x Word32 = 32 bytes)
    -- [16..31] = worm_receipt_ed25519 first 64 bytes
    scheduler_janet_array:
        bits32[32] {0,0,0,0, 0,0,0,0, 0,0,0,0,0,0,0,0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
}

-- External C/Fortran functions
foreign import ccall "sov_kv_allocate_blocks"
    sov_kv_allocate_blocks :: Word64 -> Word32 -> IO ();

foreign import ccall "sov_cuda_flash_attention"
    sov_cuda_flash_attention :: Word32 -> Word32 -> Word64 -> Word64 -> Word64 -> Word64 -> Word64 -> Word64 -> Word32 -> Word32 -> IO ();

foreign import ccall "sov_worm_checkpoint"
    sov_worm_checkpoint :: Word64 -> IO ();

foreign import ccall "sov_worm_restore"
    sov_worm_restore :: Word64 -> IO ();

foreign import ccall "sov_kv_append_tokens"
    sov_kv_append_tokens :: Word64 -> Word64 -> Word64 -> Word64 -> Word32 -> Word32 -> IO ();

foreign import ccall "sov_speculative_draft"
    sov_speculative_draft :: Word64 -> Word32 -> Word64 -> IO ();

foreign import ccall "sov_speculative_verify"
    sov_speculative_verify :: Word64 -> Word64 -> Word32 -> IO Word32;

foreign import ccall "sov_power_event"
    sov_power_event :: Word32 -> IO ();

-- =====================================================================
-- scheduler_step(state, batch_ptr, kv_ptr) -> new_state
-- =====================================================================
scheduler_step(W_ state, W_ batch_ptr, W_ kv_ptr)
{
    W_ pending, new_state, tokens_gen, power;

    power = W_[scheduler_janet_array + (5 * SIZEOF_W)];
    if (power == 1) {   -- SUSPEND
        jump scheduler_do_checkpoint(state, batch_ptr, kv_ptr);
    }
    if (power == 2) {   -- RESUME
        jump scheduler_do_resume(state, batch_ptr, kv_ptr);
    }

    switch [0..5] state {
        case 0: {   -- IDLE
            pending = W_[scheduler_janet_array + (0 * SIZEOF_W)];
            if (pending > 0) {
                new_state = 1;  -- -> PREFILL
            } else {
                new_state = 0;  -- stay IDLE
            }
            return (new_state);
        }
        case 1: {   -- PREFILL
            -- Allocate KV blocks for this batch
            foreign "C" sov_kv_allocate_blocks(kv_ptr, W_[scheduler_janet_array + (1 * SIZEOF_W)]);
            -- Reset token counter
            W_[scheduler_janet_array + (2 * SIZEOF_W)] = 0;
            new_state = 2;  -- -> GENERATE
            return (new_state);
        }
        case 2: {   -- GENERATE
            W_ heads, seqs, head_dim, block_size;
            seqs      = W_[scheduler_janet_array + (1 * SIZEOF_W)];
            heads     = 32;    -- llama-3 8B default
            head_dim  = 128;
            block_size = 16;

            -- Run flash attention forward pass
            foreign "C" sov_cuda_flash_attention(
                seqs, heads,
                batch_ptr,          -- q
                kv_ptr,             -- k
                kv_ptr + 4096,      -- v
                batch_ptr + 8192,   -- out
                kv_ptr + 16384,     -- block_table
                kv_ptr + 32768,     -- seq_lens
                head_dim, block_size);

            -- Run speculative draft
            foreign "C" sov_speculative_draft(batch_ptr, W_[scheduler_janet_array + (6 * SIZEOF_W)], kv_ptr);

            -- Increment token counter
            tokens_gen = W_[scheduler_janet_array + (2 * SIZEOF_W)] + 1;
            W_[scheduler_janet_array + (2 * SIZEOF_W)] = tokens_gen;

            -- Every 64 tokens: WORM checkpoint
            if ((tokens_gen & 63) == 0) {
                foreign "C" sov_worm_checkpoint(kv_ptr);
                -- BFT vote: increment quorum height
                W_[scheduler_janet_array + (7 * SIZEOF_W)] = W_[scheduler_janet_array + (7 * SIZEOF_W)] + 1;
            }

            -- Check if sequence complete (token_count reached max or EOS)
            if (tokens_gen >= 2048) {
                -- Decrement pending count
                W_[scheduler_janet_array + (0 * SIZEOF_W)] = W_[scheduler_janet_array + (0 * SIZEOF_W)] - 1;
                new_state = 0;  -- -> IDLE
            } else {
                new_state = 2;  -- stay GENERATE
            }
            return (new_state);
        }
        case 3: {   -- SWAP: copy KV blocks GPU->CPU
            -- Swap oldest sequence KV to CPU (simplified: just checkpoint)
            foreign "C" sov_worm_checkpoint(kv_ptr);
            new_state = 1;  -- -> PREFILL
            return (new_state);
        }
        case 4: {   -- CHECKPOINT
            jump scheduler_do_checkpoint(state, batch_ptr, kv_ptr);
        }
        case 5: {   -- RESUME
            jump scheduler_do_resume(state, batch_ptr, kv_ptr);
        }
        default: {
            new_state = 0;
            return (new_state);
        }
    }
}

scheduler_do_checkpoint(W_ state, W_ batch_ptr, W_ kv_ptr)
{
    foreign "C" sov_worm_checkpoint(kv_ptr);
    -- Receipt written to janet_array[8..31] by worm_checkpoint C impl
    W_[scheduler_janet_array + (5 * SIZEOF_W)] = 0;  -- clear power flag
    return (5);  -- -> RESUME state
}

scheduler_do_resume(W_ state, W_ batch_ptr, W_ kv_ptr)
{
    foreign "C" sov_worm_restore(kv_ptr);
    W_[scheduler_janet_array + (5 * SIZEOF_W)] = 0;  -- clear power flag
    return (1);  -- -> PREFILL
}

-- =====================================================================
-- janet_get(slot) / janet_set(slot, val)
-- =====================================================================
janet_get(W_ slot)
{
    W_ val;
    val = W_[scheduler_janet_array + (slot * SIZEOF_W)];
    return (val);
}

janet_set(W_ slot, W_ val)
{
    W_[scheduler_janet_array + (slot * SIZEOF_W)] = val;
    return ();
}