From: opencoti
Subject: [PATCH 2/N] lazy slot-context allocation — Phase 2 (grow on demand + partial zero-fill)

Phase 2 of the F4 M3 lazy slot-context rollout. Builds on Phase 1
(deferred zero-fill) by adding a soft allocation cap that
find_slot honors, a grow-on-demand path that bumps the cap when
real prompts need more cells, and a partial-range zero-fill so
RSS commits scale with actual workload — not with the configured
n_ctx_slot × slot_count product.

The end-state for the production embedder
(qwen3-0.6b, --parallel 2 --ctx-size 16384, CPU mode):
- With --slot-initial-ctx 0 (the default) Phase 2 is a no-op
  layered on Phase 1: cap == n_ctx_seq from construction; full
  buffer cleared on first batch (matches Phase 1 RSS profile).
- With --slot-initial-ctx 4096 (the recommended embedder
  setting) the warm RSS scales with whatever cells real prompts
  have touched. A run that only ever embeds 8–56-token requests
  commits ~4096 × per-cell bytes, not the full 16384.

The patch threads `slot_initial_ctx` through:
- public API: llama_context_params (include/llama.h)
- internal:   llama_cparams (src/llama-cparams.h)
- ctor chain: llama_context::ctor (src/llama-context.cpp)
              llama_kv_cache::ctor (src/llama-kv-cache.{h,cpp})
              llama_kv_cache_iswa::ctor (src/llama-kv-cache-iswa.{h,cpp})
              llama_model::create_memory call sites (src/llama-model.cpp)
- hybrid mem: src/llama-memory-hybrid.cpp + -iswa.cpp pass 0
              (kept on Phase-1 behavior — recurrent/hybrid models
              are not the F4 embedder target)
- CLI:        common_params.slot_initial_ctx + --slot-initial-ctx
              (common/common.{h,cpp}, common/arg.cpp)

Mechanism (inside llama_kv_cache):
- kv_size_max:    hard cap (== constructor's kv_size).
- kv_size_alloc:  soft cap; cells [0, kv_size_alloc) visible to
                  find_slot. Starts at min(slot_initial_ctx,
                  kv_size_max) — Phase 1 default of 0 collapses
                  to kv_size_max so find_slot sees everything.
- n_cells_cleared: half-open range of cells already zero-filled.
                  Phase 1's `bool cleared` is replaced; the cells
                  beyond n_cells_cleared have malloc'd address
                  space but uncommitted pages.
- ensure_cleared(up_to):
    Partial memset of cells [n_cells_cleared, up_to) on each
    layer's K and V tensor across every stream, using
    ggml_backend_tensor_memset (host-or-device dispatch). The
    common inference path passes up_to=0 which expands to the
    current soft cap; state_read passes up_to=kv_size_max for
    full coverage on the deserialize path.
- ensure_capacity(cells_needed):
    Geometric grow (next-pow2, n_pad-aligned) of kv_size_alloc
    up to kv_size_max. Monotonic — never shrinks (Phase 3 adds
    shrink-on-idle).
- prepare(ubatches):
    When find_slot fails inside the current soft cap, retries
    after ensure_capacity(used_max + ubatch.n_tokens) +
    ensure_cleared(). Bounded by kv_size_max; a genuine
    overflow at the hard cap propagates up unchanged.

Bench: docs/decisions/0001-lazy-slot-context.md §6.2 Phase 2 row.
Milestone: F4 M3 — see docs/features/memory_embedder.md
Design: docs/decisions/0001-lazy-slot-context.md
Upstreaming: deferred; propose to llamafile + llama.cpp
  maintainers after F4 M3's full phase-set ships and bench
  numbers are public.

---
diff --git a/llama.cpp/common/arg.cpp b/llama.cpp/common/arg.cpp
--- a/llama.cpp/common/arg.cpp
+++ b/llama.cpp/common/arg.cpp
@@ -1361,6 +1361,16 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.kv_unified = value;
         }
     ).set_env("LLAMA_ARG_KV_UNIFIED").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_PERPLEXITY, LLAMA_EXAMPLE_BATCHED, LLAMA_EXAMPLE_BENCH, LLAMA_EXAMPLE_PARALLEL}));
+    add_opt(common_arg(
+        // opencoti F4 M3 Phase 2 — see docs/decisions/0001-lazy-slot-context.md
+        {"--slot-initial-ctx"}, "N",
+        string_format("initial KV-cell allocation per sequence (default: %d, 0 = full n_ctx_seq, "
+                      "positive values cap initial allocation and enable grow-on-demand)",
+                      params.slot_initial_ctx),
+        [](common_params & params, int value) {
+            params.slot_initial_ctx = value;
+        }
+    ).set_env("LLAMA_ARG_SLOT_INITIAL_CTX").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_PARALLEL}));
     add_opt(common_arg(
         {"--cache-idle-slots"},
         {"--no-cache-idle-slots"},
diff --git a/llama.cpp/common/common.cpp b/llama.cpp/common/common.cpp
--- a/llama.cpp/common/common.cpp
+++ b/llama.cpp/common/common.cpp
@@ -1628,6 +1628,8 @@ struct llama_context_params common_context_params_to_llama(const common_params &
 
     cparams.n_ctx             = params.n_ctx;
     cparams.n_seq_max         = params.n_parallel;
+    // opencoti F4 M3 Phase 2 — see docs/decisions/0001-lazy-slot-context.md
+    cparams.slot_initial_ctx  = params.slot_initial_ctx < 0 ? 0u : (uint32_t) params.slot_initial_ctx;
     cparams.n_rs_seq          = params.speculative.need_n_rs_seq();
     cparams.n_batch           = params.n_batch;
     cparams.n_ubatch          = params.n_ubatch;
diff --git a/llama.cpp/common/common.h b/llama.cpp/common/common.h
--- a/llama.cpp/common/common.h
+++ b/llama.cpp/common/common.h
@@ -425,6 +425,11 @@ struct ggml_opt_optimizer_params common_opt_lr_pars(void * userdata);
 struct common_params {
     int32_t n_predict             =    -1; // max. number of new tokens to predict, -1 == no limit
     int32_t n_ctx                 =     0; // context size, 0 == context the model was trained with
+    // opencoti F4 M3 Phase 2 — see docs/decisions/0001-lazy-slot-context.md
+    // Initial KV-cell allocation per sequence. 0 = full n_ctx_seq allocation
+    // (Phase 1 behavior — only zero-fill is deferred). Positive values cap
+    // the initial allocation and enable grow-on-demand up to n_ctx_seq.
+    int32_t slot_initial_ctx      =     0;
     int32_t n_batch               =  2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_ubatch              =   512; // physical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep                =     0; // number of tokens to keep from initial prompt
diff --git a/llama.cpp/include/llama.h b/llama.cpp/include/llama.h
--- a/llama.cpp/include/llama.h
+++ b/llama.cpp/include/llama.h
@@ -338,6 +338,14 @@ extern "C" {
         uint32_t n_batch;           // logical maximum batch size that can be submitted to llama_decode
         uint32_t n_ubatch;          // physical maximum batch size
         uint32_t n_seq_max;         // max number of sequences (i.e. distinct states for recurrent models)
+        // opencoti F4 M3 Phase 2 — see docs/decisions/0001-lazy-slot-context.md
+        // Initial KV-cell allocation per sequence. 0 = behave like Phase 1
+        // (allocate the full n_ctx_seq up front, only zero-fill is deferred).
+        // Positive values cap the initial allocation at min(slot_initial_ctx,
+        // n_ctx_seq); subsequent prompts that exceed this cap trigger a grow
+        // up to next_pow2 (still capped at n_ctx_seq). Reduces warm-state RSS
+        // when typical prompts are much shorter than n_ctx_seq.
+        uint32_t slot_initial_ctx;
         uint32_t n_rs_seq;          // number of recurrent-state snapshots per seq for rollback (0 = no rollback) [EXPERIMENTAL]
         int32_t  n_threads;         // number of threads to use for generation
         int32_t  n_threads_batch;   // number of threads to use for batch processing
diff --git a/llama.cpp/src/llama-context.cpp b/llama.cpp/src/llama-context.cpp
--- a/llama.cpp/src/llama-context.cpp
+++ b/llama.cpp/src/llama-context.cpp
@@ -58,6 +58,9 @@ llama_context::llama_context(
         cparams.n_rs_seq = 0;
     }
 
+    // opencoti F4 M3 Phase 2 — see docs/decisions/0001-lazy-slot-context.md
+    cparams.slot_initial_ctx = params.slot_initial_ctx;
+
     cparams.n_threads        = params.n_threads;
     cparams.n_threads_batch  = params.n_threads_batch;
     cparams.yarn_ext_factor  = params.yarn_ext_factor  >= 0.0f ? params.yarn_ext_factor  : hparams.yarn_ext_factor;
@@ -3346,6 +3349,7 @@ llama_context_params llama_context_default_params() {
         /*.n_batch                     =*/ 2048,
         /*.n_ubatch                    =*/ 512,
         /*.n_seq_max                   =*/ 1,
+        /*.slot_initial_ctx            =*/ 0,
         /*.n_rs_seq                    =*/ 0,
         /*.n_threads                   =*/ GGML_DEFAULT_N_THREADS, // TODO: better default
         /*.n_threads_batch             =*/ GGML_DEFAULT_N_THREADS,
diff --git a/llama.cpp/src/llama-cparams.h b/llama.cpp/src/llama-cparams.h
--- a/llama.cpp/src/llama-cparams.h
+++ b/llama.cpp/src/llama-cparams.h
@@ -12,6 +12,10 @@ struct llama_cparams {
     uint32_t n_batch;
     uint32_t n_ubatch;
     uint32_t n_seq_max;
+    // opencoti F4 M3 Phase 2 — see docs/decisions/0001-lazy-slot-context.md
+    // 0 = full n_ctx_seq allocation. Positive value caps the initial KV-cell
+    // allocation per sequence and enables grow-on-demand up to n_ctx_seq.
+    uint32_t slot_initial_ctx;
     uint32_t n_rs_seq;        // number of recurrent-state snapshots per seq for rollback
     int32_t  n_threads;       // number of threads to use for generation
     int32_t  n_threads_batch; // number of threads to use for batch processing
diff --git a/llama.cpp/src/llama-kv-cache-iswa.cpp b/llama.cpp/src/llama-kv-cache-iswa.cpp
--- a/llama.cpp/src/llama-kv-cache-iswa.cpp
+++ b/llama.cpp/src/llama-kv-cache-iswa.cpp
@@ -20,6 +20,7 @@ llama_kv_cache_iswa::llama_kv_cache_iswa(
                      bool   swa_full,
                      bool   unified,
                  uint32_t   kv_size,
+                 uint32_t   kv_size_initial,
                  uint32_t   n_seq_max,
                  uint32_t   n_ubatch,
                  uint32_t   n_pad,
@@ -61,14 +62,14 @@ llama_kv_cache_iswa::llama_kv_cache_iswa(
 
     kv_base = std::make_unique<llama_kv_cache>(
             model, type_k, type_v,
-            v_trans, offload, unified, size_base, n_seq_max, n_pad,
+            v_trans, offload, unified, size_base, kv_size_initial, n_seq_max, n_pad,
             0, LLAMA_SWA_TYPE_NONE, filter_base, reuse);
 
     LLAMA_LOG_INFO("%s: creating     SWA KV cache, size = %u cells\n", __func__, size_swa);
 
     kv_swa = std::make_unique<llama_kv_cache>(
             model, type_k, type_v,
-            v_trans, offload, unified, size_swa, n_seq_max, n_pad,
+            v_trans, offload, unified, size_swa, kv_size_initial, n_seq_max, n_pad,
             hparams.n_swa, hparams.swa_type, filter_swa, reuse);
 }
 
diff --git a/llama.cpp/src/llama-kv-cache-iswa.h b/llama.cpp/src/llama-kv-cache-iswa.h
--- a/llama.cpp/src/llama-kv-cache-iswa.h
+++ b/llama.cpp/src/llama-kv-cache-iswa.h
@@ -22,6 +22,9 @@ public:
                          bool   swa_full,
                          bool   unified,
                      uint32_t   kv_size,
+                     // opencoti F4 M3 Phase 2 — see docs/decisions/0001-lazy-slot-context.md
+                     // Forwarded to both kv_base and kv_swa. 0 = full kv_size.
+                     uint32_t   kv_size_initial,
                      uint32_t   n_seq_max,
                      uint32_t   n_ubatch,
                      uint32_t   n_pad,
diff --git a/llama.cpp/src/llama-kv-cache.cpp b/llama.cpp/src/llama-kv-cache.cpp
--- a/llama.cpp/src/llama-kv-cache.cpp
+++ b/llama.cpp/src/llama-kv-cache.cpp
@@ -85,6 +85,7 @@ llama_kv_cache::llama_kv_cache(
                      bool   offload,
                      bool   unified,
                  uint32_t   kv_size,
+                 uint32_t   kv_size_initial,
                  uint32_t   n_seq_max,
                  uint32_t   n_pad,
                  uint32_t   n_swa,
@@ -96,6 +97,24 @@ llama_kv_cache::llama_kv_cache(
 
     GGML_ASSERT(kv_size % n_pad == 0);
 
+    // opencoti F4 M3 Phase 2 — see docs/decisions/0001-lazy-slot-context.md
+    // Initialize the hard cap and the soft cap. kv_size_initial==0 (the
+    // legacy / Phase-1 default) means "no soft cap" so the whole kv_size
+    // is visible to find_slot. A positive value is clamped to [n_pad,
+    // kv_size] and rounded up to the next n_pad multiple.
+    kv_size_max = kv_size;
+    if (kv_size_initial == 0) {
+        kv_size_alloc = kv_size_max;
+    } else {
+        uint32_t cap = std::min(kv_size_initial, kv_size_max);
+        cap = std::max(cap, n_pad);
+        cap = ((cap + n_pad - 1) / n_pad) * n_pad;
+        cap = std::min(cap, kv_size_max);
+        kv_size_alloc = cap;
+        LLAMA_LOG_INFO("%s: slot_initial_ctx = %u cells (cap = %u)\n",
+                __func__, kv_size_alloc, kv_size_max);
+    }
+
     const uint32_t n_layer_kv = hparams.n_layer_kv();
 
     // define a comparator for the buft -> ctx map to ensure that the order is well-defined:
@@ -287,7 +306,9 @@ llama_kv_cache::llama_kv_cache(
     }
 
     if (model.hparams.no_alloc) {
-        cleared = true;
+        // opencoti F4 M3 Phase 1+2 — dummy buffers are size 0; treat all
+        // cells as "already cleared" so ensure_cleared is a no-op.
+        n_cells_cleared = kv_size_max;
     } else {
         LLAMA_LOG_INFO("%s: KV cache zero-fill deferred until first batch\n", __func__);
     }
@@ -348,20 +369,81 @@ llama_kv_cache::llama_kv_cache(
     debug = LLAMA_KV_CACHE_DEBUG ? atoi(LLAMA_KV_CACHE_DEBUG) : 0;
 }
 
-// opencoti F4 M3 Phase 1 — see docs/decisions/0001-lazy-slot-context.md
-void llama_kv_cache::ensure_cleared() {
-    if (cleared) {
+// opencoti F4 M3 Phase 1+2 — see docs/decisions/0001-lazy-slot-context.md
+void llama_kv_cache::ensure_cleared(uint32_t up_to_cells) {
+    // up_to_cells == 0 means "clear up to the current soft cap". Callers
+    // that need full coverage (state_read) pass kv_size_max.
+    const uint32_t target = up_to_cells == 0
+        ? kv_size_alloc
+        : std::min(up_to_cells, kv_size_max);
+
+    if (n_cells_cleared >= target) {
         return;
     }
 
-    for (auto & [_, buf] : ctxs_bufs) {
-        LLAMA_LOG_INFO("%s: %10s KV buffer zero-fill = %8.2f MiB (deferred)\n", __func__,
-                ggml_backend_buffer_name(buf.get()),
-                ggml_backend_buffer_get_size(buf.get())/1024.0/1024.0);
-        ggml_backend_buffer_clear(buf.get(), 0);
+    const uint32_t a = n_cells_cleared;
+    const uint32_t b = target;
+
+    // Per-layer, per-stream partial memset of the K and V tensors.
+    // ggml_backend_tensor_memset handles the host/device dispatch — for the
+    // CPU backend (the only one the embedder uses today) this is a direct
+    // memset on the malloc'd buffer; for offloaded buffers it's a backend
+    // op. Either way only bytes for cells [a, b) get touched, so on Linux
+    // only those pages get committed to RSS.
+    for (const auto & layer : layers) {
+        for (uint32_t st = 0; st < n_stream; ++st) {
+            if (layer.k) {
+                const size_t off = (size_t) st * (size_t) layer.k->nb[2]
+                                 + (size_t) a  * (size_t) layer.k->nb[1];
+                const size_t sz  = (size_t) (b - a) * (size_t) layer.k->nb[1];
+                ggml_backend_tensor_memset(layer.k, 0, off, sz);
+            }
+            if (layer.v) {
+                const size_t off = (size_t) st * (size_t) layer.v->nb[2]
+                                 + (size_t) a  * (size_t) layer.v->nb[1];
+                const size_t sz  = (size_t) (b - a) * (size_t) layer.v->nb[1];
+                ggml_backend_tensor_memset(layer.v, 0, off, sz);
+            }
+        }
+    }
+
+    LLAMA_LOG_INFO("%s: cleared KV cells [%u, %u) of %u (%s)\n",
+            __func__, a, b, kv_size_max,
+            up_to_cells == 0 ? "soft cap" : "full");
+
+    n_cells_cleared = b;
+}
+
+// opencoti F4 M3 Phase 2 — see docs/decisions/0001-lazy-slot-context.md
+bool llama_kv_cache::ensure_capacity(uint32_t cells_needed) {
+    if (cells_needed <= kv_size_alloc) {
+        return true;
+    }
+
+    if (kv_size_alloc >= kv_size_max) {
+        return false;
+    }
+
+    // Grow geometrically (next power-of-two) and align up to n_pad. The
+    // pow2 ramp avoids re-growing every few tokens for long prompts; n_pad
+    // alignment preserves the constructor's GGML_ASSERT(kv_size % n_pad == 0)
+    // invariant for find_slot.
+    uint32_t new_cap = std::max<uint32_t>(kv_size_alloc, n_pad);
+    while (new_cap < cells_needed && new_cap < kv_size_max) {
+        new_cap *= 2;
     }
+    new_cap = ((new_cap + n_pad - 1) / n_pad) * n_pad;
+    new_cap = std::min(new_cap, kv_size_max);
 
-    cleared = true;
+    if (new_cap == kv_size_alloc) {
+        return cells_needed <= kv_size_alloc;
+    }
+
+    LLAMA_LOG_INFO("%s: grow soft cap %u -> %u (need %u, max %u)\n",
+            __func__, kv_size_alloc, new_cap, cells_needed, kv_size_max);
+
+    kv_size_alloc = new_cap;
+    return cells_needed <= kv_size_alloc;
 }
 
 void llama_kv_cache::clear(bool data) {
@@ -371,14 +453,16 @@ void llama_kv_cache::clear(bool data) {
     }
 
     if (data) {
-        // opencoti F4 M3 Phase 1: if the deferred zero-fill has not happened
-        // yet there is nothing to clear — v_cells.reset() above already
-        // satisfies the logical "data cleared" state. Don't commit RSS just
-        // to zero pages that have never been written.
-        if (cleared) {
-            for (auto & [_, buf] : ctxs_bufs) {
-                ggml_backend_buffer_clear(buf.get(), 0);
-            }
+        // opencoti F4 M3 Phase 1+2: if no zero-fill has happened yet there
+        // is nothing to clear — v_cells.reset() above already satisfies the
+        // logical "data cleared" state. Don't commit RSS just to zero pages
+        // that have never been written. When a partial clear is already on
+        // the books, re-issue it against the current soft cap so any cells
+        // touched between the first ensure_cleared and now are re-zeroed.
+        if (n_cells_cleared > 0) {
+            const uint32_t b = n_cells_cleared;
+            n_cells_cleared = 0;
+            ensure_cleared(b);
         }
     }
 }
@@ -743,10 +827,28 @@ llama_kv_cache::slot_info_vec_t llama_kv_cache::prepare(const std::vector<llama_
 
     for (const auto & ubatch : ubatches) {
         // only find a suitable slot for the ubatch. don't modify the cells yet
-        const auto sinfo_new = find_slot(ubatch, false);
+        auto sinfo_new = find_slot(ubatch, false);
         if (sinfo_new.empty()) {
-            success = false;
-            break;
+            // opencoti F4 M3 Phase 2 — see docs/decisions/0001-lazy-slot-context.md
+            // Phase 2 keeps the soft alloc cap (kv_size_alloc) below kv_size_max
+            // until a real ubatch needs more cells than the cap exposes. When
+            // find_slot fails, try to grow the cap to cover what we already
+            // know is used plus the ubatch's token count, then retry. The grow
+            // is bounded by kv_size_max; if we are already at the hard cap the
+            // failure is genuine and we propagate it up.
+            uint32_t used_max = 0;
+            for (uint32_t st = 0; st < n_stream; ++st) {
+                used_max = std::max(used_max, v_cells[st].used_max_p1());
+            }
+            const uint32_t want = used_max + ubatch.n_tokens;
+            if (ensure_capacity(want)) {
+                ensure_cleared();
+                sinfo_new = find_slot(ubatch, false);
+            }
+            if (sinfo_new.empty()) {
+                success = false;
+                break;
+            }
         }
 
         // remember the position that we found
@@ -981,14 +1083,20 @@ llama_kv_cache::slot_info llama_kv_cache::find_slot(const llama_ubatch & ubatch,
 
         uint32_t head_cur = v_heads[seq_to_stream[seq_id]];
 
+        // opencoti F4 M3 Phase 2 — see docs/decisions/0001-lazy-slot-context.md
+        // The Phase-1 search swept the full hard cap (cells.size() == kv_size_max).
+        // Phase 2 caps it at the current soft alloc; prepare() grows the cap and
+        // retries when the soft cap is too small for the requested ubatch.
+        const uint32_t cap = kv_size_alloc;
+
         // if we have enough unused cells before the current head ->
         //   better to start searching from the beginning of the cache, hoping to fill it
         if (head_cur > cells.get_used() + 2*n_tokens) {
             head_cur = 0;
         }
 
-        if (n_tokens > cells.size()) {
-            LLAMA_LOG_ERROR("%s: n_tokens = %d > size = %u\n", __func__, n_tokens, cells.size());
+        if (n_tokens > cap) {
+            // not necessarily fatal — prepare() may grow the cap and retry
             return { };
         }
 
@@ -999,8 +1107,8 @@ llama_kv_cache::slot_info llama_kv_cache::find_slot(const llama_ubatch & ubatch,
         const uint32_t n_test = cont ? n_tokens : 1;
 
         while (true) {
-            if (head_cur + n_test > cells.size()) {
-                n_tested += cells.size() - head_cur;
+            if (head_cur + n_test > cap) {
+                n_tested += cap - head_cur;
                 head_cur = 0;
                 continue;
             }
@@ -1058,7 +1166,7 @@ llama_kv_cache::slot_info llama_kv_cache::find_slot(const llama_ubatch & ubatch,
                 res.idxs[s].clear();
             }
 
-            if (n_tested >= cells.size()) {
+            if (n_tested >= cap) {
                 //LLAMA_LOG_ERROR("%s: failed to find a slot for %d tokens\n", __func__, n_tokens);
                 return { };
             }
@@ -1961,10 +2069,13 @@ void llama_kv_cache::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama
 
     GGML_ASSERT(seq_id == -1 || (seq_id >= 0 && (size_t) seq_id < seq_to_stream.size()));
 
-    // opencoti F4 M3 Phase 1 — see docs/decisions/0001-lazy-slot-context.md
-    // state_read writes the saved KV blob into our tensors, which requires
-    // backing storage. Force the deferred allocation before we read.
-    ensure_cleared();
+    // opencoti F4 M3 Phase 1+2 — see docs/decisions/0001-lazy-slot-context.md
+    // state_read writes the saved KV blob into our tensors at arbitrary
+    // offsets, so we cannot rely on the soft cap. Force a full-range
+    // zero-fill before we read — this commits all KV pages but only on
+    // the explicit deserialize path.
+    ensure_capacity(kv_size_max);
+    ensure_cleared(kv_size_max);
 
     uint32_t n_stream_cur;
     io.read(&n_stream_cur, sizeof(n_stream_cur));
diff --git a/llama.cpp/src/llama-kv-cache.h b/llama.cpp/src/llama-kv-cache.h
--- a/llama.cpp/src/llama-kv-cache.h
+++ b/llama.cpp/src/llama-kv-cache.h
@@ -101,6 +101,11 @@ public:
                          bool   offload,
                          bool   unified,
                      uint32_t   kv_size,
+                     // opencoti F4 M3 Phase 2 — see docs/decisions/0001-lazy-slot-context.md
+                     // Initial allocation cap. 0 means the same as kv_size
+                     // (Phase 1 behavior). Otherwise clamped to [n_pad, kv_size]
+                     // and rounded up to the next n_pad multiple.
+                     uint32_t   kv_size_initial,
                      uint32_t   n_seq_max,
                      uint32_t   n_pad,
                      uint32_t   n_swa,
@@ -152,15 +157,28 @@ public:
 
     bool get_has_shift() const;
 
-    // opencoti F4 M3 Phase 1 — see docs/decisions/0001-lazy-slot-context.md
-    // Idempotent: zeroes the KV-cache backend buffers on first call.
-    // The buffers are malloc'd eagerly at construction (so the scheduler
-    // can resolve tensor->buffer at reserve time) but NOT cleared — on
-    // Linux, malloc reserves address space without committing pages, so
-    // RSS stays low until something writes. This method does the memset
-    // that commits the pages. Called from inference entry points
-    // (init_batch, prepare, update, state_read).
-    void ensure_cleared();
+    // opencoti F4 M3 Phase 1+2 — see docs/decisions/0001-lazy-slot-context.md
+    // Idempotent. Zeroes cells [n_cells_cleared, up_to_cells) of every KV
+    // layer's K and V tensors across every stream. The buffers are malloc'd
+    // eagerly at construction (so the scheduler can resolve tensor->buffer
+    // at reserve time) but their pages are not committed to RSS until they
+    // are first written. Partial memset (via ggml_backend_tensor_memset)
+    // commits only the requested range — that is the Phase 2 RSS win.
+    //
+    // up_to_cells == 0 means "use the current soft cap kv_size_alloc"
+    // (the normal inference path). Pass kv_size_max for paths that need
+    // full coverage (state_read).
+    void ensure_cleared(uint32_t up_to_cells = 0);
+
+    // opencoti F4 M3 Phase 2 — see docs/decisions/0001-lazy-slot-context.md
+    // Grow the soft allocation cap. Returns true iff the cap covers
+    // cells_needed after the grow. Bound by kv_size_max. Idempotent at
+    // the cap.
+    bool ensure_capacity(uint32_t cells_needed);
+
+    // opencoti F4 M3 Phase 2 — see docs/decisions/0001-lazy-slot-context.md
+    // The current soft cap (cells visible to find_slot).
+    uint32_t get_kv_size_alloc() const { return kv_size_alloc; }
 
     ggml_type type_k() const;
     ggml_type type_v() const;
@@ -266,12 +284,22 @@ private:
     // ggml contexts for the KV cache along with the allocated backend buffers:
     std::vector<std::pair<ggml_context_ptr, ggml_backend_buffer_ptr>> ctxs_bufs;
 
-    // opencoti F4 M3 Phase 1 — see docs/decisions/0001-lazy-slot-context.md
-    // True once the backend buffers in ctxs_bufs have been zero-filled.
-    // The constructor sets this true for no_alloc models (whose dummy
-    // buffers are size 0 — clearing is a no-op) and false otherwise. The
-    // first ensure_cleared() call zeroes the buffers and flips this true.
-    bool cleared = false;
+    // opencoti F4 M3 Phase 1+2 — see docs/decisions/0001-lazy-slot-context.md
+    // Cells in the half-open range [0, n_cells_cleared) are guaranteed to
+    // be zeroed across every layer's K and V tensors and every stream.
+    // ensure_cleared(up_to) grows this monotonically up to kv_size_alloc
+    // (or kv_size_max for the explicit-full call). The constructor sets
+    // this to kv_size_max for no_alloc models (whose dummy buffers are
+    // size 0 — clearing is a no-op) and 0 otherwise.
+    uint32_t n_cells_cleared = 0;
+
+    // opencoti F4 M3 Phase 2 — see docs/decisions/0001-lazy-slot-context.md
+    // kv_size_max:   hard cap on cells (set from constructor's kv_size).
+    // kv_size_alloc: current soft cap. find_slot only scans cells [0, kv_size_alloc).
+    //                Grows via ensure_capacity, never shrinks (Phase 2; Phase 3
+    //                adds shrink-on-idle).
+    uint32_t kv_size_max   = 0;
+    uint32_t kv_size_alloc = 0;
 
     // the current index from where we start searching for a free slot in the ring buffer of KV cells (see find_slot())
     // note: this is not part of the KV state and it's only used to speed-up the find_slot() method
diff --git a/llama.cpp/src/llama-memory-hybrid-iswa.cpp b/llama.cpp/src/llama-memory-hybrid-iswa.cpp
--- a/llama.cpp/src/llama-memory-hybrid-iswa.cpp
+++ b/llama.cpp/src/llama-memory-hybrid-iswa.cpp
@@ -40,6 +40,10 @@ llama_memory_hybrid_iswa::llama_memory_hybrid_iswa(
         swa_full,
         unified,
         kv_size,
+        // opencoti F4 M3 Phase 2 — see docs/decisions/0001-lazy-slot-context.md
+        // Hybrid-iswa does not yet thread slot_initial_ctx; pass 0 for
+        // Phase 1 backward-compat (full eager allocation).
+        0,
         n_seq_max,
         n_ubatch,
         n_pad,
diff --git a/llama.cpp/src/llama-memory-hybrid.cpp b/llama.cpp/src/llama-memory-hybrid.cpp
--- a/llama.cpp/src/llama-memory-hybrid.cpp
+++ b/llama.cpp/src/llama-memory-hybrid.cpp
@@ -39,6 +39,10 @@ llama_memory_hybrid::llama_memory_hybrid(
         offload,
         unified,
         kv_size,
+        // opencoti F4 M3 Phase 2 — see docs/decisions/0001-lazy-slot-context.md
+        // Hybrid memory does not yet thread slot_initial_ctx; pass 0 for
+        // Phase 1 backward-compat (full eager allocation).
+        0,
         n_seq_max,
         n_pad,
         n_swa,
diff --git a/llama.cpp/src/llama-model.cpp b/llama.cpp/src/llama-model.cpp
--- a/llama.cpp/src/llama-model.cpp
+++ b/llama.cpp/src/llama-model.cpp
@@ -2073,6 +2073,8 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                                 params.swa_full,
                                 cparams.kv_unified,
                                 cparams.n_ctx_seq,
+                                // opencoti F4 M3 Phase 2 — see docs/decisions/0001-lazy-slot-context.md
+                                cparams.slot_initial_ctx,
                                 cparams.n_seq_max,
                                 cparams.n_ubatch,
                                 1,
@@ -2089,6 +2091,8 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                                 cparams.offload_kqv,
                                 cparams.kv_unified,
                                 cparams.n_ctx_seq,
+                                // opencoti F4 M3 Phase 2 — see docs/decisions/0001-lazy-slot-context.md
+                                cparams.slot_initial_ctx,
                                 cparams.n_seq_max,
                                 1,
                                 hparams.n_swa,