prefill: STRATA_PREFILL_CPU_SHARE_BATCH=1 (opt-in) gives the prompt path the CPU pool beside batch slots, for a read with no slot decoding

The share was off whenever --batch was set (share_pool = batch <= 0). With the variable set the pool goes to the
prompt path as without batch slots, and before each read the serve loop sets Prefill::set_share_blocked from the
slots: a read with a slot decoding takes no share (the slots' windows use the pool), one with none does.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

include/strata/prefill/prefill.hpp

@@ -245,6 +245,9 @@ public:
     /// `by_default`: this path is one the share was measured on (CUDA, one GPU, no batch slots), so with the variable unset
     /// the share is on for chunks below 1024 tokens (STRATA_PREFILL_CPU_SHARE=0 turns it off).
     static void arm_cpu_share(bool applies, bool by_default = false);
+    /// STRATA_PREFILL_CPU_SHARE_BATCH=1 (batch slots beside the pool): set before each read - true while a slot decodes
+    /// (its windows use the pool), so no chunk of that read takes the share; false lets the read's chunks take it.
+    static void set_share_blocked(bool blocked);
 
 private:
     static uint64_t bytes_needed_impl(const core::ModelGeometry& g, const core::SessionState& ss, int64_t chunk,

src/prefill/prefill.cpp

@@ -226,6 +226,8 @@ inline int64_t cpu_share_max() {
 // On a layer split every stage may have the pool (set_cpu_pool), and the stages read different chunks at the same time:
 // one stage at a time takes it for a chunk, the other reads its chunk with the GPU alone.
 std::atomic<bool> g_cpu_pool_busy{false};
+// STRATA_PREFILL_CPU_SHARE_BATCH: a slot is decoding beside this read (Prefill::set_share_blocked): the pool is its windows'
+std::atomic<bool> g_share_blocked{false};
 double g_pinned_share = 1.0;
 // The streamed ring: 384 slots when (nearly) every streamed expert is DMA'd from pinned RAM - measured on Q2_0,
 // 8192-token chunks: 96 slots 1153 tok/s, 384 1294 (the next layer's experts arrive during its attention half) -
@@ -2385,6 +2387,7 @@ void Prefill::arm_cpu_share(bool applies, bool by_default) {
     if (applies && cpu_share_on()) g_stream_min_share = cpu_share_max();
 }
 
+void Prefill::set_share_blocked(bool blocked) { g_share_blocked.store(blocked, std::memory_order_release); }
 bool Prefill::ring_bytes_enabled() { return ring_bytes_on(); }
 size_t Prefill::slot_guard_bytes() { return mmq::direct_slot_guard(); }
 
@@ -2849,7 +2852,8 @@ bool Prefill::run_impl(const int64_t* tokens, int64_t n, int64_t pos0, std::stri
             bool held = false;
             ~PoolHold() { if (held) g_cpu_pool_busy.store(false, std::memory_order_release); }
         } pool_hold;
-        if (cpu_pool_ != nullptr && cpu_share_on() && !stream_all && !pipe_chunked) {
+        if (cpu_pool_ != nullptr && cpu_share_on() && !stream_all && !pipe_chunked &&
+            !g_share_blocked.load(std::memory_order_acquire)) {
             bool expect = false;
             pool_hold.held = g_cpu_pool_busy.compare_exchange_strong(expect, true, std::memory_order_acq_rel);
         }

src/program/generate.cpp

@@ -8108,7 +8108,13 @@ int main(int argc, char** argv) {
             }
         }
 #endif
-        const bool share_pool = o.batch <= 0 && !o.no_pool;
+        // STRATA_PREFILL_CPU_SHARE_BATCH=1 (opt-in): the prompt path gets the pool beside batch slots too; a read takes the
+        // share only while no slot decodes (set_share_blocked before each read), since the slots' windows use the pool
+        static const bool share_batch = [] { const char* v = std::getenv("STRATA_PREFILL_CPU_SHARE_BATCH"); return v != nullptr && v[0] == '1'; }();
+        const bool share_pool = (o.batch <= 0 || share_batch) && !o.no_pool;
+        if (share_batch && o.batch > 0 && !o.no_pool)
+            std::fprintf(stderr, "strata serve: STRATA_PREFILL_CPU_SHARE_BATCH=1: the prompt path may use the CPU pool beside the "
+                                 "%lld batch slots, for a read with no slot decoding\n", (long long) o.batch);
         bool share_has_work = strata::prefill::share_rules::default_share_has_work(
             (int64_t) xcache.slots(), (int64_t) (g.n_layers * g.n_expert));   // #1595
         if (!share_has_work)
@@ -13068,6 +13074,19 @@ int main(int argc, char** argv) {
             // the request ends with `YIELDED <slot> <tokens>` + DONE cancel, and the server sends it again later: it
             // continues from the slot with the same chunks.  #656's cooperative preemption, with a slot as the park.
             int64_t req_chunk = 0, req_chunk_len = 0;   // the DONE line's chunk=: the longest batched segment's first chunk
+
+            // STRATA_PREFILL_CPU_SHARE_BATCH: the read takes the CPU share only with no slot decoding (their windows use the pool)
+            auto share_gate = [&]() {
+                int busy = 0;
+                for (const auto& x : bs) busy += x.active ? 1 : 0;
+                strata::prefill::Prefill::set_share_blocked(busy > 0);
+                static int said = 0;
+                static const bool on = [] { const char* v = std::getenv("STRATA_PREFILL_CPU_SHARE_BATCH"); return v != nullptr && v[0] == '1'; }();
+                if (on && said < 4) {
+                    ++said;
+                    std::fprintf(stderr, "strata batch: CPU share for this read: %s (%d slot(s) decoding)\n", busy > 0 ? "off" : "on", busy);
+                }
+            };
             auto read_part = [&](int64_t a0, int64_t b0, std::string& e) -> bool {
                 // --prefill-pipe on a layer split: the segment's chunk is planned ONCE, on its whole length (or taken
                 // from the read it goes on from: resume_plan), and every run of it reads in that chunk (force_chunk) -
@@ -13115,6 +13134,7 @@ int main(int argc, char** argv) {
                 } seg_plan_scope{seg_plan};
                 auto run = [&](int64_t q, int64_t r) -> bool {
                     if (planned) sp.force_chunk(plan.chunk, plan_single, plan.chunked);
+                    share_gate();
                     return sp.run(ids.data() + q, r - q, q, e);
                 };
                 // (a layer split reads its stages as a pipeline over one run's chunks: beside slots a chunk at a time, with
@@ -13339,6 +13359,7 @@ int main(int argc, char** argv) {
                         std::string read_err;
                         in_prompt_read = true;
                         sp.stop_at_chunk = [&] { return want_yield.load(std::memory_order_acquire) >= 0; };
+                        share_gate();   // on this thread, before the reader starts
                         std::thread reader([&] {
                             conc_thread_cpus();
                             try {   // (an exception leaving a std::thread ends the process: a failed read instead)