Add DSpark speculative draft runtime

.gitignore

@@ -6,6 +6,7 @@
 /ds4_native
 /ds4_server_test
 /ds4_test
+/tests/test_q4k_dot
 /ds4flash.gguf
 /TODO.md
 /gguf/
@@ -16,3 +17,4 @@ __pycache__/
 /misc/
 .*.swp
 .DS_Store
+/logs/

AGENT.md

@@ -28,6 +28,24 @@ Objective-C only where Metal requires it and Metal kernels under `metal/`.
 - Avoid large CPU inference runs on macOS; the CPU path has previously exposed kernel VM failures with very large mappings.
 - Do not run multiple huge model processes concurrently. The instance lock is intentional.
 
+
+## Repository Maintenance
+
+- In this checkout, `origin` is the audreyt/ds4 fork. To chase antirez
+  upstream, fetch it explicitly:
+  `git fetch https://github.com/antirez/ds4.git main:refs/remotes/antirez/main`.
+- Compare and merge `antirez/main`; do not merge `origin/main` expecting
+  upstream changes.
+- Leave upstream-chase merge commits unpushed unless the user asks.
+- Preserve local README/MODEL_CARD benchmark numbers unless replaced by fresh
+  local measurements.
+- Treat scheduling, KV-cache lifetime, attention math, tokenizer behavior,
+  model shape, and tensor metadata conflicts as correctness-sensitive.
+- Keep CUDA/ROCm parity in view when upstream changes Metal logic.
+- Historical branch note: older M5 side experiments used Metal function
+  constant slot 703 after avoiding slot 702. Verify current branches before
+  reusing those slots.
+
 ## Layout
 
 - `ds4.c`: model loading, tokenizer, CPU reference code, Metal graph scheduling,

CONTRIBUTING.md

@@ -129,6 +129,17 @@ and background load when comparing two commits. For backend work, run at least
 one before/after CSV and compare both `prefill_tps` and `gen_tps`. Generation is
 greedy and skips EOS so each frontier gets the same number of generated tokens.
 
+When comparing this fork against antirez/main or another fork:
+
+- Use each fork's preferred quant only when the question is fork-vs-fork
+  behavior; use the same GGUF when isolating runtime or kernel changes.
+- Run baseline and candidate sequentially on the same machine. Do not run two
+  huge model processes concurrently.
+- Use the current sweep above unless the PR notes explicitly choose another
+  sweep. Do not copy old README tables or stale agent playbooks.
+- Report exact commits, model files, CSV paths, backend, hardware, and thermal
+  state.
+
 To generate a graph for a CSV:
 
 ```sh

Makefile

@@ -13,12 +13,13 @@ OBJCFLAGS ?= -O3 -ffast-math $(DEBUG_FLAGS) $(NATIVE_CPU_FLAG) -Wall -Wextra -fo
 
 LDLIBS ?= -lm -pthread
 METAL_SRCS := $(wildcard metal/*.metal)
+CUDA_CONFIG := .ds4_cuda.config
 ROCM_SRCS := $(wildcard rocm/*.cuh)
 
 ifeq ($(UNAME_S),Darwin)
 METAL_LDLIBS := $(LDLIBS) -framework Foundation -framework Metal
-CORE_OBJS = ds4.o ds4_distributed.o ds4_ssd.o ds4_metal.o
-CPU_CORE_OBJS = ds4_cpu.o ds4_distributed.o ds4_ssd.o
+CORE_OBJS = ds4.o ds4_distributed.o ds4_ssd.o ds4_dspark_runtime.o ds4_metal.o
+CPU_CORE_OBJS = ds4_cpu.o ds4_distributed.o ds4_ssd.o ds4_dspark_runtime.o
 else
 CFLAGS += -D_GNU_SOURCE -fno-finite-math-only
 CUDA_HOME ?= /usr/local/cuda
@@ -28,8 +29,8 @@ ifneq ($(strip $(CUDA_ARCH)),)
 NVCC_ARCH_FLAGS := -arch=$(CUDA_ARCH)
 endif
 NVCCFLAGS ?= -O3 -g -lineinfo --use_fast_math $(NVCC_ARCH_FLAGS) -Xcompiler $(NATIVE_CPU_FLAG) -Xcompiler -pthread
-CORE_OBJS = ds4.o ds4_distributed.o ds4_ssd.o ds4_cuda.o
-CPU_CORE_OBJS = ds4_cpu.o ds4_distributed.o ds4_ssd.o
+CORE_OBJS = ds4.o ds4_distributed.o ds4_ssd.o ds4_dspark_runtime.o ds4_cuda.o
+CPU_CORE_OBJS = ds4_cpu.o ds4_distributed.o ds4_ssd.o ds4_dspark_runtime.o
 CUDA_LDLIBS ?= -lm -Xcompiler -pthread -L$(CUDA_HOME)/targets/sbsa-linux/lib -L$(CUDA_HOME)/lib64 -lcudart -lcublas
 HIPCC ?= $(shell command -v hipcc 2>/dev/null || echo /opt/rocm/bin/hipcc)
 ROCM_ARCH ?= gfx1151
@@ -40,7 +41,7 @@ DS4_LINK_LIBS ?= $(CUDA_LDLIBS)
 METAL_LDLIBS := $(LDLIBS)
 endif
 
-.PHONY: all help clean test cpu cuda cuda-spark cuda-generic cuda-regression strix-halo rocm
+.PHONY: all help clean test cpu cuda cuda-spark cuda-generic cuda-regression strix-halo rocm FORCE
 
 ifeq ($(UNAME_S),Darwin)
 all: ds4 ds4-server ds4-bench ds4-eval ds4-agent
@@ -106,7 +107,7 @@ cuda:
 
 strix-halo:
 	$(MAKE) -B ds4 ds4-server ds4-bench ds4-eval ds4-agent \
-		CORE_OBJS="ds4.o ds4_distributed.o ds4_ssd.o ds4_rocm.o" \
+		CORE_OBJS="ds4.o ds4_distributed.o ds4_ssd.o ds4_dspark_runtime.o ds4_rocm.o" \
 		CFLAGS="$(CFLAGS) -DDS4_ROCM_BUILD" \
 		DS4_LINK="$(HIPCC) $(ROCM_CFLAGS)" \
 		DS4_LINK_LIBS="$(ROCM_LDLIBS)"
@@ -137,13 +138,29 @@ cpu: ds4_cli_cpu.o ds4_server_cpu.o ds4_bench_cpu.o ds4_eval_cpu.o ds4_agent_cpu
 
 cuda-regression: tests/cuda_long_context_smoke
 	./tests/cuda_long_context_smoke
+
+$(CUDA_CONFIG): FORCE
+	@tmp="$@.tmp"; \
+	{ \
+		printf '%s\n' "CUDA_ARCH=$(CUDA_ARCH)"; \
+		printf '%s\n' "NVCC=$(NVCC)"; \
+		printf '%s\n' "NVCCFLAGS=$(NVCCFLAGS)"; \
+	} > "$$tmp"; \
+	if test -r "$@" && cmp -s "$$tmp" "$@"; then \
+		rm -f "$$tmp"; \
+	else \
+		mv "$$tmp" "$@"; \
+		rm -f ds4_cuda.o; \
+	fi
 endif
 
-ds4.o: ds4.c ds4.h ds4_ssd.h ds4_distributed.h ds4_gpu.h
+ds4.o: ds4.c ds4.h ds4_ssd.h ds4_distributed.h ds4_dspark_runtime.h ds4_gpu.h
 	$(CC) $(CFLAGS) -c -o $@ ds4.c
 
 ds4_ssd.o: ds4_ssd.c ds4_ssd.h
-	$(CC) $(CFLAGS) -c -o $@ ds4_ssd.c
+
+ds4_dspark_runtime.o: ds4_dspark_runtime.c ds4_dspark_runtime.h ds4.h
+	$(CC) $(CFLAGS) -c -o $@ ds4_dspark_runtime.c
 
 ds4_cli.o: ds4_cli.c ds4.h ds4_ssd.h ds4_distributed.h ds4_help.h linenoise.h
 	$(CC) $(CFLAGS) -c -o $@ ds4_cli.c
@@ -187,7 +204,7 @@ rax.o: rax.c rax.h rax_malloc.h
 linenoise.o: linenoise.c linenoise.h
 	$(CC) $(CFLAGS) -c -o $@ linenoise.c
 
-ds4_cpu.o: ds4.c ds4.h ds4_ssd.h ds4_distributed.h ds4_gpu.h
+ds4_cpu.o: ds4.c ds4.h ds4_ssd.h ds4_distributed.h ds4_dspark_runtime.h ds4_gpu.h
 	$(CC) $(CFLAGS) -DDS4_NO_GPU -c -o $@ ds4.c
 
 ds4_cli_cpu.o: ds4_cli.c ds4.h ds4_ssd.h ds4_distributed.h ds4_help.h linenoise.h
@@ -208,7 +225,7 @@ ds4_agent_cpu.o: ds4_agent.c ds4.h ds4_ssd.h ds4_distributed.h ds4_help.h ds4_kv
 ds4_metal.o: ds4_metal.m ds4_gpu.h $(METAL_SRCS)
 	$(CC) $(OBJCFLAGS) -c -o $@ ds4_metal.m
 
-ds4_cuda.o: ds4_cuda.cu ds4_gpu.h ds4_iq2_tables_cuda.inc
+ds4_cuda.o: ds4_cuda.cu ds4_gpu.h ds4_iq2_tables_cuda.inc $(CUDA_CONFIG)
 	$(NVCC) $(NVCCFLAGS) -c -o $@ ds4_cuda.cu
 
 ds4_rocm.o: ds4_rocm.cu ds4_gpu.h ds4_iq2_tables_cuda.inc $(ROCM_SRCS)
@@ -241,4 +258,4 @@ q4k-dot-test: tests/test_q4k_dot.c
 	./tests/test_q4k_dot
 
 clean:
-	rm -f ds4 ds4-server ds4-bench ds4-eval ds4-agent ds4_cpu ds4_native ds4_server_test ds4_test ds4_agent_test tests/test_q4k_dot *.o tests/cuda_long_context_smoke tests/cuda_long_context_smoke.o
+	rm -f ds4 ds4-server ds4-bench ds4-eval ds4-agent ds4_cpu ds4_native ds4_server_test ds4_test ds4_agent_test tests/test_q4k_dot *.o tests/cuda_long_context_smoke tests/cuda_long_context_smoke.o $(CUDA_CONFIG) $(CUDA_CONFIG).tmp