[remoting] Rebase on top of b6945 #15
Conversation
|
Caution Review failedFailed to post review comments WalkthroughThis PR introduces a comprehensive GGML remoting backend and frontend infrastructure, enabling distributed tensor computation across virtgpu. Adds dispatched backend handling, RPC serialization protocols, frontend buffer/device management, DRM kernel interfaces, and build/run scripts. Disables verbose logging in llama.cpp and adds timing instrumentation. Changes
Sequence Diagram(s)sequenceDiagram
participant Frontend as GGML Frontend
participant Dispatcher as Backend Dispatcher
participant GPU as Remote GPU
participant Host as Host Virtgpu
Frontend->>Host: create_virtgpu() - handshake & load backend library
Host-->>Frontend: virtgpu handle, capset, shmem regions
Note over Frontend: Graph Compute Flow
Frontend->>Frontend: serialize_graph(cgraph)
Frontend->>Host: remote_call_prepare(GRAPH_COMPUTE)
Host-->>Frontend: encoder + decoder
Frontend->>Frontend: encode cgraph to shmem
Frontend->>Host: remote_call() - send command
Host->>Dispatcher: apir_backend_dispatcher(cmd_type, encoded_data)
Dispatcher->>GPU: backend_graph_compute(cgraph)
GPU-->>Dispatcher: ggml_status result
Dispatcher->>Host: encode status
Host-->>Frontend: remote_call result
Frontend->>Frontend: deserialize result
Frontend->>Host: remote_call_finish()
sequenceDiagram
participant App as Application
participant FrontBuf as Frontend Buffer
participant RemoteCall as Remote Backend
participant GPU as GPU Device
participant SharedMem as Shared Memory
App->>FrontBuf: set_tensor(tensor, data)
FrontBuf->>SharedMem: allocate/use shmem region
FrontBuf->>FrontBuf: encode tensor + shmem_id
FrontBuf->>RemoteCall: apir_buffer_set_tensor(encoded)
RemoteCall->>SharedMem: resolve shmem pointer
RemoteCall->>GPU: dma/copy data to GPU buffer
GPU-->>RemoteCall: done
RemoteCall-->>FrontBuf: status
FrontBuf-->>App: set_tensor complete
Estimated code review effort🎯 5 (Critical) | ⏱️ ~90+ minutes Justification: This is a substantial architectural addition with high complexity and heterogeneity:
Areas requiring extra attention:
Possibly related PRs
Suggested reviewers
Poem
Pre-merge checks and finishing touches❌ Failed checks (1 warning)
✅ Passed checks (1 passed)
✨ Finishing touches
🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
|
[APPROVALNOTIFIER] This PR is NOT APPROVED This pull-request has been approved by: The full list of commands accepted by this bot can be found here.
Needs approval from an approver in each of these files:
Approvers can indicate their approval by writing |
|
/test topsail |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 51
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (1)
tools/run/run.cpp (1)
1175-1199: Callstop_timer()on every exit path.We start the timer at the top of each loop iteration, but exit paths (decode failure, EOG break, token-to-string failure) return or break before the trailing
stop_timer(). Those iterations never incrementtimer_total/timer_count, so the stats under-report work and may leave the laststart_timer()unmatched. Stop the timer before each early exit.- if (llama_decode(llama_data.context.get(), batch)) { + if (llama_decode(llama_data.context.get(), batch)) { + stop_timer(); printe("failed to decode\n"); return 1; } // sample the next token, check is it an end of generation? new_token_id = llama_sampler_sample(llama_data.sampler.get(), llama_data.context.get(), -1); - if (llama_vocab_is_eog(vocab, new_token_id)) { + if (llama_vocab_is_eog(vocab, new_token_id)) { + stop_timer(); break; } std::string piece; - if (convert_token_to_string(vocab, new_token_id, piece)) { + if (convert_token_to_string(vocab, new_token_id, piece)) { + stop_timer(); return 1; } print_word_and_concatenate_to_response(piece, response); // prepare the next batch with the sampled token batch = llama_batch_get_one(&new_token_id, 1); stop_timer();
🧹 Nitpick comments (19)
ggml/src/ggml-remotingfrontend/include/venus_hw.h (1)
30-33: Consider documenting the version fields.The first four fields lack inline comments explaining their purpose and valid value ranges. While the names are descriptive, brief documentation would improve maintainability, especially for protocol version compatibility checks.
Example:
struct virgl_renderer_capset_venus { + /* Wire protocol format version for Venus commands */ uint32_t wire_format_version; + /* Vulkan XML specification version */ uint32_t vk_xml_version; + /* VK_EXT_command_serialization specification version */ uint32_t vk_ext_command_serialization_spec_version; + /* VK_MESA_venus_protocol specification version */ uint32_t vk_mesa_venus_protocol_spec_version;ggml/src/ggml-remotingfrontend/virtgpu-utils.h (2)
83-94: Consider adding [[noreturn]] attribute to FATAL.Since FATAL always calls
abort(), it would be beneficial to mark it with the[[noreturn]]attribute to help the compiler with optimization and dead code analysis.Apply this diff:
+[[noreturn]] inline void FATAL(const char *format, ...) {
96-100: Consider removing redundant wrapper function.
util_is_power_of_two_nonzero64is a simple wrapper around theIS_POT_NONZEROmacro with no additional logic. Consider using the macro directly or documenting why the function wrapper is necessary.src/llama-model.cpp (1)
2256-2256: Keep debug logs; they’re already gated by LLAMA_LOG_DEBUG.Commenting them out removes useful diagnostics for device assignment/splits. Since these are debug-level, they won’t spam unless enabled. Recommend restoring them.
- //LLAMA_LOG_DEBUG("load_tensors: layer %3d assigned to device %s, is_swa = %d\n", il, ggml_backend_dev_name(cpu_dev), is_swa); + LLAMA_LOG_DEBUG("load_tensors: layer %3d assigned to device %s, is_swa = %d\n", il, ggml_backend_dev_name(cpu_dev), is_swa); @@ - //LLAMA_LOG_DEBUG("load_tensors: layer %3d assigned to device %s, is_swa = %d\n", il, ggml_backend_dev_name(dev), is_swa); + LLAMA_LOG_DEBUG("load_tensors: layer %3d assigned to device %s, is_swa = %d\n", il, ggml_backend_dev_name(dev), is_swa);Also applies to: 2262-2262
ggml/src/ggml-remotingfrontend/include/drm-uapi/drm.h (1)
1-64: Document the kernel source and version of this vendored UAPI header.This appears to be a DRM UAPI header copied from the Linux kernel. It's important to document:
- Which kernel version this header is from
- The update policy for keeping it in sync with kernel changes
- Why vendoring is preferred over using system headers
Consider adding a comment at the top indicating this is from kernel UAPI and the specific version/commit.
ggml/src/ggml-remotingfrontend/include/drm-uapi/virtgpu_drm.h (1)
1-38: Document kernel version and note dependency on drm.h.Similar to drm.h, this vendored virtgpu UAPI header should document its kernel source version. Additionally, this header depends on drm.h which has a missing drm_mode.h include (see drm.h review comments).
run.vulkan.sh (2)
12-12: Remove commented-out code.Dead code should be removed rather than left commented out. Use version control to recover it if needed.
Apply this diff:
-#rm -f /usr/lib64/libvulkan_virtio.so -
16-16: Make MESA_FLAVOR configurable.The
MESA_FLAVORvariable is hardcoded to "good". Consider making this configurable via environment variable or command-line argument to support different testing scenarios.Apply this diff:
-MESA_FLAVOR=good +MESA_FLAVOR="${MESA_FLAVOR:-good}"ggml/src/ggml-remotingfrontend/virtgpu-forward-impl.h (1)
7-8: Clarify or remove incomplete CACHED macro.The
CACHEDmacro is defined as empty with a commented-out placeholder. If this is for future use, consider adding a TODO comment explaining its purpose. If it's not needed, remove it.src/llama-vocab.cpp (3)
2360-2361: Debug log suppression is fine; prefer a feature toggle.Commenting out LLAMA_LOG_DEBUG reduces diagnosability. Consider guarding with a compile-time flag or env-driven verbosity check instead of hard-disabling.
3207-3207: Early-return disables print_info entirely.This short-circuits all vocab diagnostics. Recommend gating via a runtime flag (e.g., LLAMA_QUIET=1) or compile-time option so developers can re-enable when needed.
3573-3573: Same as above: print_info() is now a no-op.Apply the same gating approach to allow opt-in diagnostics.
prepare.vulkan.sh (1)
1-7: Missing shebang and strict mode; add for robustness.Shellcheck SC2148 applies. Also suggest set -euo pipefail.
Apply this diff:
+#!/usr/bin/env bash +set -euo pipefail cmake -S . \ -B ../build.vulkan \ -DGGML_VULKAN=ON \ -DGGML_NATIVE=OFF \ -DGGML_METAL=OFF \ -DLLAMA_CURL=OFF \ -DCMAKE_BUILD_TYPE=Debugsrc/llama-model-loader.cpp (2)
682-709: Log lines disabled; prefer a controllable verbosity switch.Commenting out LLAMA_LOG_INFO/DEBUG reduces helpful diagnostics. Gate via env var (e.g., LLAMA_VERBOSE_META) or compile-time option instead of hard-disabling.
Also applies to: 793-794
1160-1161: print_info() early-return suppresses all loader metadata.Consider honoring a quiet flag rather than unconditional return, to aid debugging when needed.
build.backend.sh (1)
13-13: Consider separating declaration and assignment to catch errors.Combining
exportwith command substitution can mask failures from the subcommand.Apply this diff to separate the operations:
-export SDKROOT=$(xcrun --sdk macosx --show-sdk-path) +SDKROOT=$(xcrun --sdk macosx --show-sdk-path) +export SDKROOTpodman_compile.sh (1)
34-34: Consider mounting only the required directory instead of entire $HOME.Mounting the entire home directory gives the container broad access to user data. For better security isolation, mount only the workspace directory needed for the build.
For example, if only the current project directory is needed:
--env HOME="$HOME" \ --env PERF_MODE="${PERF_MODE:-}" \ --env BENCH_MODE="${BENCH_MODE:-}" \ --v "$HOME":"$HOME":Z \ +-v "$PWD":"$PWD":Z \ -w "$PWD" \ggml/src/ggml-remotingfrontend/CMakeLists.txt (1)
27-29: Usepkg-configorfind_packageinstead of hardcoding system paths.The hardcoded path
/usr/include/libdrm/and the Fedora-specificdnf installcomment reduce portability. Different Linux distributions and macOS may install libdrm in different locations.Apply this diff to use pkg-config for better portability:
# dnf install -y libdrm-devel -target_link_libraries(ggml-remotingfrontend PUBLIC drm) -target_include_directories(ggml-remotingfrontend PUBLIC /usr/include/libdrm/) +find_package(PkgConfig REQUIRED) +pkg_check_modules(DRM REQUIRED libdrm) +target_link_libraries(ggml-remotingfrontend PUBLIC ${DRM_LIBRARIES}) +target_include_directories(ggml-remotingfrontend PUBLIC ${DRM_INCLUDE_DIRS})ggml/src/ggml-remotingbackend/backend-dispatched.h (1)
56-65: Fix command name strings to match handlers.Several command-name strings drop the
_device/_backendprefixes (e.g., Line 56 returns"backend_get_device_count"while the handler isbackend_reg_get_device_count). When this helper is used for tracing or diagnostics, the mismatches make it very hard to map logs back to the actual dispatchers. Please align the returned strings with the real handler names.Apply this diff:
- case APIR_COMMAND_TYPE_DEVICE_GET_COUNT: return "backend_get_device_count"; - case APIR_COMMAND_TYPE_DEVICE_GET_NAME: return "backend_get_device_name"; - case APIR_COMMAND_TYPE_DEVICE_GET_DESCRIPTION: return "backend_get_device_description"; - case APIR_COMMAND_TYPE_DEVICE_GET_MEMORY: return "backend_get_device_memory"; - case APIR_COMMAND_TYPE_DEVICE_GET_BUFFER_TYPE: return "backend_get_buffer_type"; - case APIR_COMMAND_TYPE_DEVICE_GET_PROPS: return "backend_get_props"; - case APIR_COMMAND_TYPE_DEVICE_BUFFER_FROM_PTR: return "backend_buffer_from_ptr"; + case APIR_COMMAND_TYPE_DEVICE_GET_COUNT: return "backend_reg_get_device_count"; + case APIR_COMMAND_TYPE_DEVICE_GET_NAME: return "backend_device_get_name"; + case APIR_COMMAND_TYPE_DEVICE_GET_DESCRIPTION: return "backend_device_get_description"; + case APIR_COMMAND_TYPE_DEVICE_GET_MEMORY: return "backend_device_get_memory"; + case APIR_COMMAND_TYPE_DEVICE_GET_BUFFER_TYPE: return "backend_device_get_buffer_type"; + case APIR_COMMAND_TYPE_DEVICE_GET_PROPS: return "backend_device_get_props"; + case APIR_COMMAND_TYPE_DEVICE_BUFFER_FROM_PTR: return "backend_device_buffer_from_ptr";
📜 Review details
Configuration used: CodeRabbit UI
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (76)
.gitignore(1 hunks)CMakePresets.json(1 hunks)OWNERS(1 hunks)build.backend.sh(1 hunks)build.remoting.sh(1 hunks)build.sh(1 hunks)build.vulkan.sh(1 hunks)ggml/CMakeLists.txt(2 hunks)ggml/include/ggml-remoting-frontend.h(1 hunks)ggml/src/CMakeLists.txt(1 hunks)ggml/src/ggml-backend-reg.cpp(3 hunks)ggml/src/ggml-metal/CMakeLists.txt(1 hunks)ggml/src/ggml-metal/ggml-metal-context.m(1 hunks)ggml/src/ggml-metal/ggml-metal-device.m(2 hunks)ggml/src/ggml-metal/ggml-metal-remoting.cpp(1 hunks)ggml/src/ggml-remotingbackend/CMakeLists.txt(1 hunks)ggml/src/ggml-remotingbackend/backend-convert.h(1 hunks)ggml/src/ggml-remotingbackend/backend-dispatched-backend.cpp(1 hunks)ggml/src/ggml-remotingbackend/backend-dispatched-buffer-type.cpp(1 hunks)ggml/src/ggml-remotingbackend/backend-dispatched-buffer.cpp(1 hunks)ggml/src/ggml-remotingbackend/backend-dispatched-device.cpp(1 hunks)ggml/src/ggml-remotingbackend/backend-dispatched-metal.cpp(1 hunks)ggml/src/ggml-remotingbackend/backend-dispatched.cpp(1 hunks)ggml/src/ggml-remotingbackend/backend-dispatched.h(1 hunks)ggml/src/ggml-remotingbackend/backend-internal.h(1 hunks)ggml/src/ggml-remotingbackend/backend-utils.h(1 hunks)ggml/src/ggml-remotingbackend/backend.cpp(1 hunks)ggml/src/ggml-remotingbackend/shared/api_remoting.h(1 hunks)ggml/src/ggml-remotingbackend/shared/apir_backend.h(1 hunks)ggml/src/ggml-remotingbackend/shared/venus_cs.h(1 hunks)ggml/src/ggml-remotingbackend/shared/venus_cs_ggml-rpc.cpp(1 hunks)ggml/src/ggml-remotingbackend/shared/venus_cs_ggml-rpc.h(1 hunks)ggml/src/ggml-remotingbackend/shared/venus_cs_ggml.h(1 hunks)ggml/src/ggml-remotingbackend/venus_cs_ggml-rpc-back.cpp(1 hunks)ggml/src/ggml-remotingfrontend/CMakeLists.txt(1 hunks)ggml/src/ggml-remotingfrontend/ggml-backend-buffer-type.cpp(1 hunks)ggml/src/ggml-remotingfrontend/ggml-backend-buffer.cpp(1 hunks)ggml/src/ggml-remotingfrontend/ggml-backend-device.cpp(1 hunks)ggml/src/ggml-remotingfrontend/ggml-backend-host-buffer-type.cpp(1 hunks)ggml/src/ggml-remotingfrontend/ggml-backend-reg.cpp(1 hunks)ggml/src/ggml-remotingfrontend/ggml-backend.cpp(1 hunks)ggml/src/ggml-remotingfrontend/ggml-metal-remoting.cpp(1 hunks)ggml/src/ggml-remotingfrontend/ggml-metal-remoting.h(1 hunks)ggml/src/ggml-remotingfrontend/ggml-remoting-frontend.cpp(1 hunks)ggml/src/ggml-remotingfrontend/ggml-remoting.h(1 hunks)ggml/src/ggml-remotingfrontend/include/drm-uapi/drm.h(1 hunks)ggml/src/ggml-remotingfrontend/include/drm-uapi/virtgpu_drm.h(1 hunks)ggml/src/ggml-remotingfrontend/include/venus_hw.h(1 hunks)ggml/src/ggml-remotingfrontend/venus_cs_ggml-rpc-front.cpp(1 hunks)ggml/src/ggml-remotingfrontend/virtgpu-forward-backend.cpp(1 hunks)ggml/src/ggml-remotingfrontend/virtgpu-forward-buffer-type.cpp(1 hunks)ggml/src/ggml-remotingfrontend/virtgpu-forward-buffer.cpp(1 hunks)ggml/src/ggml-remotingfrontend/virtgpu-forward-device.cpp(1 hunks)ggml/src/ggml-remotingfrontend/virtgpu-forward-impl.h(1 hunks)ggml/src/ggml-remotingfrontend/virtgpu-forward-metal.cpp(1 hunks)ggml/src/ggml-remotingfrontend/virtgpu-forward.h(1 hunks)ggml/src/ggml-remotingfrontend/virtgpu-shm.cpp(1 hunks)ggml/src/ggml-remotingfrontend/virtgpu-shm.h(1 hunks)ggml/src/ggml-remotingfrontend/virtgpu-utils.cpp(1 hunks)ggml/src/ggml-remotingfrontend/virtgpu-utils.h(1 hunks)ggml/src/ggml-remotingfrontend/virtgpu.cpp(1 hunks)ggml/src/ggml-remotingfrontend/virtgpu.h(1 hunks)podman_compile.sh(1 hunks)prepare.backend.sh(1 hunks)prepare.remoting.sh(1 hunks)prepare.sh(1 hunks)prepare.vulkan.sh(1 hunks)run.ramalama.sh(1 hunks)run.remoting.sh(1 hunks)run.vulkan.sh(1 hunks)src/llama-context.cpp(2 hunks)src/llama-kv-cache.cpp(1 hunks)src/llama-model-loader.cpp(5 hunks)src/llama-model.cpp(2 hunks)src/llama-vocab.cpp(3 hunks)tools/run/run.cpp(3 hunks)
🧰 Additional context used
🧬 Code graph analysis (42)
ggml/include/ggml-remoting-frontend.h (1)
ggml/src/ggml-remotingfrontend/ggml-backend-reg.cpp (2)
ggml_backend_remoting_frontend_reg(132-159)ggml_backend_remoting_frontend_reg(132-132)
ggml/src/ggml-backend-reg.cpp (1)
ggml/src/ggml-remotingfrontend/ggml-backend-reg.cpp (2)
ggml_backend_remoting_frontend_reg(132-159)ggml_backend_remoting_frontend_reg(132-132)
ggml/src/ggml-remotingfrontend/virtgpu-forward-metal.cpp (2)
ggml/src/ggml-remotingbackend/shared/venus_cs.h (1)
vn_decode_bool_t(508-512)ggml/src/ggml-remotingfrontend/virtgpu.cpp (2)
remote_call_finish(551-567)remote_call_finish(552-555)
ggml/src/ggml-remotingfrontend/ggml-metal-remoting.h (1)
ggml/src/ggml-remotingfrontend/ggml-metal-remoting.cpp (4)
get_metal_dev_context(4-18)get_metal_dev_context(4-4)ggml_metal_device_supports_op(20-254)ggml_metal_device_supports_op(20-20)
ggml/src/ggml-remotingbackend/backend-dispatched-metal.cpp (2)
ggml/src/ggml-remotingbackend/backend-utils.h (1)
ERROR(52-55)ggml/src/ggml-remotingbackend/shared/venus_cs.h (1)
vn_encode_bool_t(502-506)
ggml/src/ggml-remotingbackend/backend-convert.h (1)
ggml/src/ggml-remotingfrontend/ggml-remoting.h (2)
ggml_buffer_to_apir_handle(137-139)ggml_buffer_type_to_apir_handle(34-38)
ggml/src/ggml-remotingfrontend/virtgpu-forward-buffer-type.cpp (4)
ggml/src/ggml-remotingbackend/shared/venus_cs_ggml.h (1)
vn_encode_ggml_buffer_type(69-73)ggml/src/ggml-remotingbackend/shared/venus_cs.h (7)
vn_decode_array_size_unchecked(293-299)vn_cs_decoder_alloc_array(493-498)vn_decode_char_array(466-476)vn_decode_size_t(394-400)vn_decode_bool_t(508-512)vn_encode_size_t(387-392)vn_decode_apir_buffer_host_handle_t(536-540)ggml/src/ggml-remotingfrontend/virtgpu-utils.h (3)
FATAL(83-94)INFO(35-44)INFO(46-47)ggml/src/ggml-remotingfrontend/virtgpu.cpp (2)
remote_call_finish(551-567)remote_call_finish(552-555)
ggml/src/ggml-remotingfrontend/venus_cs_ggml-rpc-front.cpp (1)
ggml/src/ggml-remotingfrontend/virtgpu-utils.h (1)
FATAL(83-94)
ggml/src/ggml-remotingfrontend/ggml-backend-buffer.cpp (2)
ggml/src/ggml-remotingfrontend/virtgpu-forward-buffer.cpp (12)
apir_buffer_get_base(3-21)apir_buffer_get_base(4-4)apir_buffer_set_tensor(23-65)apir_buffer_set_tensor(24-25)apir_buffer_get_tensor(68-76)apir_buffer_get_tensor(69-70)apir_buffer_get_tensor(78-114)apir_buffer_get_tensor(79-80)apir_buffer_clear(117-132)apir_buffer_clear(118-119)apir_buffer_free_buffer(135-148)apir_buffer_free_buffer(136-136)ggml/src/ggml-remotingbackend/shared/apir_backend.h (2)
start_timer(91-95)stop_timer(98-108)
ggml/src/ggml-remotingfrontend/virtgpu-shm.cpp (3)
ggml/src/ggml-remotingfrontend/virtgpu-utils.h (3)
align64(102-107)INFO(35-44)INFO(46-47)ggml/src/ggml-remotingfrontend/virtgpu.h (1)
virtgpu_ioctl(101-105)ggml/src/ggml-remotingfrontend/virtgpu-utils.cpp (2)
util_sparse_array_get(125-186)util_sparse_array_get(126-126)
ggml/src/ggml-remotingfrontend/virtgpu-shm.h (1)
ggml/src/ggml-remotingfrontend/virtgpu-shm.cpp (4)
virtgpu_shmem_create(79-111)virtgpu_shmem_create(80-80)virtgpu_shmem_destroy(71-77)virtgpu_shmem_destroy(72-73)
ggml/src/ggml-remotingbackend/backend-dispatched.cpp (3)
ggml/src/ggml-backend-reg.cpp (4)
reg(241-254)reg(241-241)reg(309-332)reg(309-309)ggml/src/ggml-remotingfrontend/virtgpu-utils.h (4)
FATAL(83-94)ERROR(72-81)INFO(35-44)INFO(46-47)ggml/src/ggml-remotingbackend/backend-utils.h (2)
ERROR(52-55)INFO(42-45)
ggml/src/ggml-remotingbackend/backend-utils.h (1)
ggml/src/ggml-remotingfrontend/virtgpu-utils.h (5)
INFO(35-44)INFO(46-47)WARNING(61-70)ERROR(72-81)FATAL(83-94)
tools/run/run.cpp (1)
src/llama-batch.cpp (2)
llama_batch_get_one(851-863)llama_batch_get_one(851-853)
ggml/src/ggml-remotingbackend/shared/venus_cs_ggml.h (6)
ggml/src/ggml-remotingbackend/shared/venus_cs.h (7)
vn_encode(136-142)vn_cs_decoder_use_inplace(81-92)vn_cs_encoder_write(108-123)vn_cs_decoder_read(98-106)vn_encode_uint32_t(342-346)vn_decode_uint32_t(348-352)vn_decode_uint64_t_array_inplace(211-215)ggml/src/ggml-remotingbackend/shared/venus_cs_ggml-rpc.cpp (8)
serialize_tensor(17-46)serialize_tensor(18-18)deserialize_tensor(48-78)deserialize_tensor(49-49)serialize_graph(96-117)serialize_graph(97-97)deserialize_graph(144-167)deserialize_graph(145-145)ggml/src/ggml.c (2)
ggml_tensor_overhead(1356-1358)ggml_init(1487-1527)ggml/src/ggml-remotingbackend/venus_cs_ggml-rpc-back.cpp (4)
deserialize_tensor(33-68)deserialize_tensor(34-34)deserialize_graph(95-118)deserialize_graph(96-96)ggml/src/ggml-remotingbackend/backend-convert.h (2)
ggml_buffer_type_to_apir_handle(11-15)ggml_buffer_to_apir_handle(5-9)ggml/src/ggml-remotingfrontend/virtgpu-utils.h (2)
FATAL(83-94)WARNING(61-70)
ggml/src/ggml-remotingfrontend/virtgpu.h (1)
ggml/src/ggml-remotingfrontend/virtgpu.cpp (10)
vn_log(347-361)vn_log(348-348)create_virtgpu(185-237)create_virtgpu(186-186)remote_call_prepare(510-549)remote_call_prepare(511-514)remote_call(569-668)remote_call(570-575)remote_call_finish(551-567)remote_call_finish(552-555)
ggml/src/ggml-remotingfrontend/ggml-backend.cpp (3)
ggml/src/ggml-remotingbackend/shared/apir_backend.h (2)
start_timer(91-95)stop_timer(98-108)ggml/src/ggml-remotingfrontend/virtgpu-forward-backend.cpp (2)
apir_backend_graph_compute(9-54)apir_backend_graph_compute(10-10)ggml/src/ggml-remotingfrontend/ggml-backend-reg.cpp (2)
ggml_backend_remoting_frontend_reg(132-159)ggml_backend_remoting_frontend_reg(132-132)
ggml/src/ggml-remotingbackend/venus_cs_ggml-rpc-back.cpp (2)
ggml/src/ggml.c (7)
ggml_new_tensor_4d(1740-1749)ggml_nbytes(1203-1226)ggml_set_name(1808-1815)ggml_tensor_overhead(1356-1358)ggml_graph_overhead_custom(6700-6702)ggml_init(1487-1527)ggml_new_graph_custom(6708-6750)ggml/src/ggml-remotingbackend/shared/venus_cs_ggml-rpc.cpp (4)
deserialize_tensor(48-78)deserialize_tensor(49-49)create_node(119-142)create_node(120-123)
ggml/src/ggml-remotingfrontend/virtgpu-forward-backend.cpp (5)
ggml/src/ggml-remotingbackend/shared/venus_cs_ggml.h (4)
vn_serialize_ggml_cgraph(140-145)vn_encode_virtgpu_shmem_res_id(128-131)vn_encode_cgraph_data(147-152)vn_decode_ggml_status(121-124)ggml/src/ggml-remotingfrontend/virtgpu-shm.cpp (4)
virtgpu_shmem_create(79-111)virtgpu_shmem_create(80-80)virtgpu_shmem_destroy(71-77)virtgpu_shmem_destroy(72-73)ggml/src/ggml-remotingfrontend/virtgpu-utils.h (2)
WARNING(61-70)FATAL(83-94)ggml/src/ggml-remotingbackend/shared/venus_cs.h (2)
vn_encode_size_t(387-392)vn_cs_new_encoder(37-46)ggml/src/ggml-remotingfrontend/virtgpu.cpp (2)
remote_call_finish(551-567)remote_call_finish(552-555)
ggml/src/ggml-remotingfrontend/ggml-backend-device.cpp (3)
ggml/src/ggml-remotingfrontend/virtgpu-forward-device.cpp (16)
apir_device_get_name(27-53)apir_device_get_name(28-28)apir_device_get_description(55-77)apir_device_get_description(56-56)apir_device_get_type(79-102)apir_device_get_type(80-80)apir_device_get_memory(104-138)apir_device_get_memory(105-105)apir_device_supports_op(140-158)apir_device_supports_op(141-141)apir_device_get_props(178-201)apir_device_get_props(179-183)apir_device_get_buffer_type(160-176)apir_device_get_buffer_type(161-161)apir_device_buffer_from_ptr(203-237)apir_device_buffer_from_ptr(204-206)ggml/src/ggml-remotingfrontend/ggml-metal-remoting.cpp (2)
ggml_metal_device_supports_op(20-254)ggml_metal_device_supports_op(20-20)ggml/src/ggml-remotingfrontend/ggml-backend.cpp (2)
ggml_backend_remoting_device_init(73-87)ggml_backend_remoting_device_init(73-73)
ggml/src/ggml-remotingfrontend/virtgpu-forward-buffer.cpp (5)
ggml/src/ggml-remotingbackend/shared/venus_cs.h (4)
vn_encode_apir_buffer_host_handle_t(530-534)vn_decode_uintptr_t(550-554)vn_encode_size_t(387-392)vn_encode_uint8_t(150-154)ggml/src/ggml-remotingfrontend/virtgpu.cpp (2)
remote_call_finish(551-567)remote_call_finish(552-555)ggml/src/ggml-remotingfrontend/virtgpu-utils.h (4)
INFO(35-44)INFO(46-47)FATAL(83-94)WARNING(61-70)ggml/src/ggml-remotingbackend/shared/venus_cs_ggml.h (2)
vn_encode_ggml_tensor(39-44)vn_encode_virtgpu_shmem_res_id(128-131)ggml/src/ggml-remotingfrontend/virtgpu-shm.cpp (4)
virtgpu_shmem_create(79-111)virtgpu_shmem_create(80-80)virtgpu_shmem_destroy(71-77)virtgpu_shmem_destroy(72-73)
ggml/src/ggml-remotingbackend/backend-dispatched.h (6)
ggml/src/ggml-remotingbackend/backend-dispatched.cpp (2)
backend_dispatch_initialize(19-47)backend_dispatch_initialize(19-19)ggml/src/ggml-remotingbackend/backend-dispatched-device.cpp (18)
backend_reg_get_device_count(9-18)backend_reg_get_device_count(9-9)backend_device_get_name(20-31)backend_device_get_name(20-20)backend_device_get_description(33-45)backend_device_get_description(34-34)backend_device_get_type(47-56)backend_device_get_type(48-48)backend_device_get_memory(58-70)backend_device_get_memory(59-59)backend_device_supports_op(72-83)backend_device_supports_op(73-73)backend_device_get_buffer_type(85-95)backend_device_get_buffer_type(86-86)backend_device_get_props(97-111)backend_device_get_props(98-98)backend_device_buffer_from_ptr(113-142)backend_device_buffer_from_ptr(114-114)ggml/src/ggml-remotingbackend/backend-dispatched-buffer-type.cpp (10)
backend_buffer_type_get_name(9-22)backend_buffer_type_get_name(10-10)backend_buffer_type_get_alignment(24-34)backend_buffer_type_get_alignment(25-25)backend_buffer_type_get_max_size(36-46)backend_buffer_type_get_max_size(37-37)backend_buffer_type_is_host(48-58)backend_buffer_type_is_host(49-49)backend_buffer_type_alloc_buffer(60-81)backend_buffer_type_alloc_buffer(61-61)ggml/src/ggml-remotingbackend/backend-dispatched-buffer.cpp (10)
backend_buffer_get_base(12-22)backend_buffer_get_base(13-13)backend_buffer_set_tensor(24-71)backend_buffer_set_tensor(25-25)backend_buffer_get_tensor(73-109)backend_buffer_get_tensor(74-74)backend_buffer_clear(111-125)backend_buffer_clear(112-112)backend_buffer_free_buffer(127-143)backend_buffer_free_buffer(128-128)ggml/src/ggml-remotingbackend/backend-dispatched-backend.cpp (2)
backend_graph_compute(13-58)backend_graph_compute(14-14)ggml/src/ggml-remotingbackend/backend-dispatched-metal.cpp (2)
backend_metal_get_device_context(14-41)backend_metal_get_device_context(15-15)
ggml/src/ggml-remotingfrontend/ggml-remoting.h (4)
ggml/src/ggml-remotingbackend/backend-convert.h (2)
ggml_buffer_type_to_apir_handle(11-15)ggml_buffer_to_apir_handle(5-9)ggml/src/ggml-remotingfrontend/ggml-backend-reg.cpp (2)
ggml_backend_remoting_get_device(49-52)ggml_backend_remoting_get_device(49-49)ggml/src/ggml-remotingfrontend/ggml-backend.cpp (2)
ggml_backend_remoting_device_init(73-87)ggml_backend_remoting_device_init(73-73)ggml/src/ggml-remotingfrontend/ggml-backend-device.cpp (2)
ggml_backend_remoting_device_get_buffer_type(129-144)ggml_backend_remoting_device_get_buffer_type(130-130)
ggml/src/ggml-remotingfrontend/virtgpu-forward-device.cpp (4)
ggml/src/ggml-remotingbackend/shared/venus_cs.h (10)
vn_decode_int32_t(235-239)vn_decode_array_size_unchecked(293-299)vn_cs_decoder_alloc_array(493-498)vn_decode_char_array(466-476)vn_decode_uint32_t(348-352)vn_decode_size_t(394-400)vn_decode_bool_t(508-512)vn_decode_apir_buffer_type_host_handle_t(522-526)vn_encode_size_t(387-392)vn_decode_apir_buffer_host_handle_t(536-540)ggml/src/ggml-remotingfrontend/virtgpu-utils.h (3)
INFO(35-44)INFO(46-47)FATAL(83-94)ggml/src/ggml-remotingfrontend/virtgpu.cpp (2)
remote_call_finish(551-567)remote_call_finish(552-555)ggml/src/ggml-remotingfrontend/virtgpu-shm.cpp (2)
virtgpu_shmem_create(79-111)virtgpu_shmem_create(80-80)
ggml/src/ggml-remotingbackend/backend-dispatched-device.cpp (3)
ggml/src/ggml-remotingbackend/shared/venus_cs.h (7)
vn_encode_int32_t(229-233)vn_encode_array_size(274-278)vn_encode_char_array(459-464)vn_encode_uint32_t(342-346)vn_encode_size_t(387-392)vn_encode_bool_t(502-506)vn_decode_size_t(394-400)ggml/src/ggml-remotingbackend/shared/venus_cs_ggml.h (4)
vn_decode_ggml_tensor_inplace(212-236)vn_encode_ggml_buffer_type(69-73)vn_decode_virtgpu_shmem_res_id(133-136)vn_encode_ggml_buffer(98-102)ggml/src/ggml-remotingfrontend/virtgpu-utils.h (1)
FATAL(83-94)
ggml/src/ggml-remotingfrontend/virtgpu-forward.h (5)
ggml/src/ggml-remotingfrontend/virtgpu-forward-device.cpp (18)
apir_device_get_count(3-25)apir_device_get_count(4-4)apir_device_get_name(27-53)apir_device_get_name(28-28)apir_device_get_description(55-77)apir_device_get_description(56-56)apir_device_get_type(79-102)apir_device_get_type(80-80)apir_device_get_memory(104-138)apir_device_get_memory(105-105)apir_device_supports_op(140-158)apir_device_supports_op(141-141)apir_device_get_buffer_type(160-176)apir_device_get_buffer_type(161-161)apir_device_get_props(178-201)apir_device_get_props(179-183)apir_device_buffer_from_ptr(203-237)apir_device_buffer_from_ptr(204-206)ggml/src/ggml-remotingfrontend/virtgpu-forward-buffer-type.cpp (10)
apir_buffer_type_get_name(3-29)apir_buffer_type_get_name(4-4)apir_buffer_type_get_alignment(31-51)apir_buffer_type_get_alignment(32-32)apir_buffer_type_get_max_size(53-73)apir_buffer_type_get_max_size(54-54)apir_buffer_type_is_host(75-95)apir_buffer_type_is_host(76-76)apir_buffer_type_alloc_buffer(97-119)apir_buffer_type_alloc_buffer(98-98)ggml/src/ggml-remotingfrontend/virtgpu-forward-buffer.cpp (12)
apir_buffer_get_base(3-21)apir_buffer_get_base(4-4)apir_buffer_set_tensor(23-65)apir_buffer_set_tensor(24-25)apir_buffer_get_tensor(68-76)apir_buffer_get_tensor(69-70)apir_buffer_get_tensor(78-114)apir_buffer_get_tensor(79-80)apir_buffer_clear(117-132)apir_buffer_clear(118-119)apir_buffer_free_buffer(135-148)apir_buffer_free_buffer(136-136)ggml/src/ggml-remotingfrontend/virtgpu-forward-backend.cpp (2)
apir_backend_graph_compute(9-54)apir_backend_graph_compute(10-10)ggml/src/ggml-remotingfrontend/virtgpu-forward-metal.cpp (2)
apir_metal_get_device_context(3-20)apir_metal_get_device_context(4-4)
ggml/src/ggml-remotingfrontend/ggml-metal-remoting.cpp (3)
ggml/src/ggml-remotingfrontend/virtgpu-forward-metal.cpp (2)
apir_metal_get_device_context(3-20)apir_metal_get_device_context(4-4)ggml/src/ggml.c (5)
ggml_get_unary_op(1794-1797)ggml_is_contiguous(1386-1388)ggml_get_glu_op(1799-1802)ggml_is_contiguous_1(1394-1396)ggml_is_contiguous_rows(1419-1423)ggml/src/ggml-impl.h (1)
ggml_get_op_params_i32(151-154)
ggml/src/ggml-remotingbackend/backend-dispatched-buffer.cpp (5)
ggml/src/ggml-remotingbackend/shared/venus_cs_ggml.h (3)
vn_decode_ggml_buffer(104-112)vn_decode_ggml_tensor(46-236)vn_decode_virtgpu_shmem_res_id(133-136)ggml/src/ggml-remotingbackend/shared/venus_cs.h (3)
vn_encode_uintptr_t(544-548)vn_decode_size_t(394-400)vn_decode_uint8_t(156-160)ggml/src/ggml-remotingbackend/shared/apir_backend.h (2)
start_timer(91-95)stop_timer(98-108)ggml/src/ggml-remotingfrontend/virtgpu-utils.h (2)
FATAL(83-94)WARNING(61-70)ggml/src/ggml-remotingbackend/venus_cs_ggml-rpc-back.cpp (2)
untrack_backend_buffer(17-26)untrack_backend_buffer(18-18)
ggml/src/ggml-remotingbackend/backend-internal.h (2)
ggml/src/ggml-remotingbackend/backend.cpp (6)
apir_backend_initialize(49-114)apir_backend_initialize(49-49)apir_backend_deinit(23-47)apir_backend_deinit(23-23)apir_backend_dispatcher(116-150)apir_backend_dispatcher(116-119)ggml/src/ggml-remotingbackend/backend-dispatched-metal.cpp (1)
ggml_backend_metal_get_device_context_fct(9-12)
ggml/src/ggml-remotingbackend/backend-dispatched-buffer-type.cpp (3)
ggml/src/ggml-remotingbackend/shared/venus_cs_ggml.h (2)
vn_decode_ggml_buffer_type(75-82)vn_encode_ggml_buffer(98-102)ggml/src/ggml-remotingbackend/shared/venus_cs.h (5)
vn_encode_array_size(274-278)vn_encode_char_array(459-464)vn_encode_size_t(387-392)vn_encode_bool_t(502-506)vn_decode_size_t(394-400)ggml/src/ggml-remotingbackend/shared/venus_cs_ggml-rpc.cpp (2)
track_backend_buffer(12-15)track_backend_buffer(13-13)
ggml/src/ggml-remotingfrontend/ggml-backend-host-buffer-type.cpp (3)
ggml/src/ggml-remotingfrontend/virtgpu-utils.h (4)
WARNING(61-70)FATAL(83-94)INFO(35-44)INFO(46-47)ggml/src/ggml-remotingfrontend/virtgpu-shm.cpp (2)
virtgpu_shmem_destroy(71-77)virtgpu_shmem_destroy(72-73)ggml/src/ggml-remotingfrontend/virtgpu-forward-device.cpp (2)
apir_device_buffer_from_ptr(203-237)apir_device_buffer_from_ptr(204-206)
ggml/src/ggml-remotingbackend/shared/venus_cs_ggml-rpc.h (3)
ggml/src/ggml-remotingbackend/shared/venus_cs_ggml-rpc.cpp (14)
serialize_tensor(17-46)serialize_tensor(18-18)serialize_graph(96-117)serialize_graph(97-97)track_backend_buffer(12-15)track_backend_buffer(13-13)add_tensor(80-94)add_tensor(81-81)deserialize_tensor(48-78)deserialize_tensor(49-49)create_node(119-142)create_node(120-123)deserialize_graph(144-167)deserialize_graph(145-145)ggml/src/ggml-remotingfrontend/venus_cs_ggml-rpc-front.cpp (6)
serialize_tensor(12-48)serialize_tensor(13-13)serialize_graph(66-87)serialize_graph(67-67)add_tensor(50-64)add_tensor(51-51)ggml/src/ggml-remotingbackend/venus_cs_ggml-rpc-back.cpp (12)
track_backend_buffer(12-15)track_backend_buffer(13-13)untrack_backend_buffer(17-26)untrack_backend_buffer(18-18)get_track_backend_buffers(28-31)get_track_backend_buffers(29-29)deserialize_tensor(33-68)deserialize_tensor(34-34)create_node(70-93)create_node(71-74)deserialize_graph(95-118)deserialize_graph(96-96)
ggml/src/ggml-remotingbackend/backend-dispatched-backend.cpp (5)
ggml/src/ggml-remotingbackend/shared/apir_backend.h (2)
start_timer(91-95)stop_timer(98-108)ggml/src/ggml-remotingbackend/shared/venus_cs_ggml.h (3)
vn_decode_virtgpu_shmem_res_id(133-136)vn_decode_ggml_cgraph(154-167)vn_encode_ggml_status(116-119)ggml/src/ggml-remotingfrontend/virtgpu-utils.h (2)
FATAL(83-94)ERROR(72-81)ggml/src/ggml-remotingbackend/shared/venus_cs.h (2)
vn_decode_size_t(394-400)vn_cs_new_decoder(27-35)ggml/src/ggml.c (2)
ggml_graph_node(6885-6893)ggml_op_desc(1273-1283)
ggml/src/ggml-remotingfrontend/virtgpu-utils.h (2)
ggml/src/ggml-remotingfrontend/virtgpu-utils.cpp (8)
thks_bye(189-195)thks_bye(189-189)breakpoint(197-200)breakpoint(197-197)util_sparse_array_get(125-186)util_sparse_array_get(126-126)util_sparse_array_init(33-41)util_sparse_array_init(34-35)ggml/src/ggml-remotingbackend/backend-utils.h (3)
INFO(42-45)WARNING(47-50)ERROR(52-55)
ggml/src/ggml-remotingfrontend/virtgpu.cpp (6)
ggml/src/ggml-remotingfrontend/virtgpu-utils.h (7)
FATAL(83-94)INFO(35-44)INFO(46-47)ERROR(72-81)WARNING(61-70)MESSAGE(50-59)os_time_sleep(126-133)ggml/src/ggml-remotingbackend/shared/venus_cs.h (3)
vn_encode_uint32_t(342-346)vn_decode_uint32_t(348-352)vn_encode_int32_t(229-233)ggml/src/ggml-remotingbackend/shared/apir_backend.h (3)
apir_backend_initialize_error(123-139)start_timer(91-95)stop_timer(98-108)ggml/src/ggml-remotingfrontend/virtgpu-utils.cpp (2)
util_sparse_array_init(33-41)util_sparse_array_init(34-35)ggml/src/ggml-remotingfrontend/virtgpu-shm.cpp (2)
virtgpu_shmem_create(79-111)virtgpu_shmem_create(80-80)ggml/src/ggml-remotingfrontend/virtgpu.h (2)
vn_log(55-98)virtgpu_ioctl(101-105)
ggml/src/ggml-remotingfrontend/ggml-backend-buffer-type.cpp (3)
ggml/src/ggml-remotingfrontend/virtgpu-utils.h (1)
FATAL(83-94)ggml/src/ggml-remotingfrontend/virtgpu-forward-device.cpp (2)
apir_device_buffer_from_ptr(203-237)apir_device_buffer_from_ptr(204-206)ggml/src/ggml-remotingfrontend/virtgpu-forward-buffer-type.cpp (8)
apir_buffer_type_get_name(3-29)apir_buffer_type_get_name(4-4)apir_buffer_type_get_alignment(31-51)apir_buffer_type_get_alignment(32-32)apir_buffer_type_get_max_size(53-73)apir_buffer_type_get_max_size(54-54)apir_buffer_type_is_host(75-95)apir_buffer_type_is_host(76-76)
ggml/src/ggml-remotingfrontend/ggml-backend-reg.cpp (4)
ggml/src/ggml-remotingfrontend/virtgpu.cpp (2)
create_virtgpu(185-237)create_virtgpu(186-186)ggml/src/ggml-remotingfrontend/virtgpu-forward-device.cpp (2)
apir_device_get_count(3-25)apir_device_get_count(4-4)ggml/src/ggml-remotingfrontend/ggml-metal-remoting.cpp (2)
get_metal_dev_context(4-18)get_metal_dev_context(4-4)ggml/src/ggml-remotingbackend/shared/apir_backend.h (1)
show_timer(110-121)
ggml/src/ggml-remotingbackend/backend.cpp (6)
ggml/src/ggml-remotingbackend/venus_cs_ggml-rpc-back.cpp (4)
get_track_backend_buffers(28-31)get_track_backend_buffers(29-29)untrack_backend_buffer(17-26)untrack_backend_buffer(18-18)ggml/src/ggml-remotingbackend/backend-utils.h (2)
INFO(42-45)ERROR(52-55)ggml/src/ggml-remotingbackend/shared/apir_backend.h (1)
show_timer(110-121)ggml/src/ggml-remotingbackend/backend-dispatched.cpp (3)
ggml_backend_reg_fct(23-23)backend_dispatch_initialize(19-47)backend_dispatch_initialize(19-19)ggml/src/ggml-remotingbackend/backend-dispatched-metal.cpp (1)
ggml_backend_metal_get_device_context_fct(9-12)ggml/src/ggml-remotingbackend/backend-dispatched.h (1)
backend_dispatch_command_name(52-88)
ggml/src/ggml-remotingbackend/shared/venus_cs_ggml-rpc.cpp (3)
ggml/src/ggml.c (7)
ggml_new_tensor_4d(1740-1749)ggml_nbytes(1203-1226)ggml_set_name(1808-1815)ggml_tensor_overhead(1356-1358)ggml_graph_overhead_custom(6700-6702)ggml_init(1487-1527)ggml_new_graph_custom(6708-6750)ggml/src/ggml-remotingfrontend/venus_cs_ggml-rpc-front.cpp (2)
serialize_tensor(12-48)serialize_tensor(13-13)ggml/src/ggml-remotingbackend/venus_cs_ggml-rpc-back.cpp (2)
deserialize_tensor(33-68)deserialize_tensor(34-34)
ggml/src/ggml-remotingbackend/shared/venus_cs.h (1)
ggml/src/ggml-remotingfrontend/virtgpu-utils.h (1)
FATAL(83-94)
ggml/src/ggml-remotingbackend/shared/apir_backend.h (2)
ggml/src/ggml-remotingbackend/backend-utils.h (1)
INFO(42-45)ggml/src/ggml-remotingfrontend/virtgpu-utils.h (2)
INFO(35-44)INFO(46-47)
ggml/src/ggml-remotingfrontend/virtgpu-utils.cpp (1)
ggml/src/ggml-remotingfrontend/virtgpu-utils.h (2)
INFO(35-44)INFO(46-47)
🪛 Shellcheck (0.11.0)
prepare.backend.sh
[error] 1-1: Tips depend on target shell and yours is unknown. Add a shebang or a 'shell' directive.
(SC2148)
build.vulkan.sh
[error] 1-1: Tips depend on target shell and yours is unknown. Add a shebang or a 'shell' directive.
(SC2148)
build.sh
[error] 1-1: Tips depend on target shell and yours is unknown. Add a shebang or a 'shell' directive.
(SC2148)
build.remoting.sh
[error] 1-1: Tips depend on target shell and yours is unknown. Add a shebang or a 'shell' directive.
(SC2148)
build.backend.sh
[error] 1-1: Tips depend on target shell and yours is unknown. Add a shebang or a 'shell' directive.
(SC2148)
[warning] 13-13: Declare and assign separately to avoid masking return values.
(SC2155)
prepare.remoting.sh
[error] 1-1: Tips depend on target shell and yours is unknown. Add a shebang or a 'shell' directive.
(SC2148)
prepare.vulkan.sh
[error] 1-1: Tips depend on target shell and yours is unknown. Add a shebang or a 'shell' directive.
(SC2148)
prepare.sh
[error] 1-1: Tips depend on target shell and yours is unknown. Add a shebang or a 'shell' directive.
(SC2148)
run.ramalama.sh
[error] 1-1: Tips depend on target shell and yours is unknown. Add a shebang or a 'shell' directive.
(SC2148)
🔇 Additional comments (25)
ggml/src/ggml-remotingfrontend/include/venus_hw.h (1)
1-22: I need to gather more information to verify the vendoring and attribution. Let me try alternative approaches:Now let me check the full file content in the repository:
Based on my verification, the copyright and attribution appear legitimate. Venus is a Virtio-GPU protocol for Vulkan command serialization, with protocol definition hosted at venus-protocol and the renderer at virglrenderer. This code originates from the Chromium/ChromiumOS ecosystem, where it has been developed as part of established virtualization infrastructure.
The MIT-style license is appropriate for permissive open-source work. The attribution is clear and complete in the header.
ggml/src/ggml-remotingfrontend/virtgpu-utils.h (3)
1-29: LGTM!The header setup, branch prediction macros, and bit manipulation utilities follow standard patterns and are correctly implemented.
34-48: LGTM!The conditional compilation of the INFO macro based on NDEBUG is a standard pattern for debug logging that optimizes away in release builds.
126-133: Verify error handling is sufficient for your use case.The function correctly handles EINTR by looping with the remaining time. However, other error codes from
clock_nanosleep(such as EINVAL or EFAULT) are silently ignored. Consider whether these errors should be logged or handled.If error handling is needed, consider:
inline void os_time_sleep(int64_t usecs) { struct timespec time; time.tv_sec = usecs / 1000000; time.tv_nsec = (usecs % 1000000) * 1000; - while (clock_nanosleep(CLOCK_MONOTONIC, 0, &time, &time) == EINTR); + int ret; + while ((ret = clock_nanosleep(CLOCK_MONOTONIC, 0, &time, &time)) == EINTR); + if (ret != 0) { + ERROR("clock_nanosleep failed with error: %d", ret); + } }ggml/src/ggml-remotingfrontend/virtgpu-utils.cpp (4)
1-5: LGTM!The includes are appropriate for the utility implementations in this file.
33-41: LGTM!The initialization correctly zeros the structure, sets the element size, and validates that node_size is a power of two >= 2.
53-186: LGTM!The sparse array implementation correctly handles:
- Thread-safe lazy allocation using atomic compare-and-swap operations
- Dynamic tree growth when indices exceed current capacity
- Race conditions by freeing the losing allocation in
_util_sparse_array_set_or_free_node- Level-based traversal for efficient sparse storage
The comments clearly explain the design choices, particularly around single-node growth for correctness.
197-200: LGTM!The
breakpoint()function serves as a convenient debug hook. The inline INFO call and comment make the intent clear.src/llama-model.cpp (1)
6447-6447: Early returns in print_info() functions are correct and necessary; the review comment's suggestion is based on a misconception.The logging infrastructure does not automatically filter by log level in the default case. While callbacks can implement level-based filtering (as seen in
common/common.cppandtools/run/run.cpp), the default callback (llama_log_callback_default) ignores the level parameter and always prints to stderr. The early returns are the only mechanism preventing unconditional verbose output during model loading. Removing them without implementing a filtering callback would cause all model info to print unconditionally, breaking default behavior for CLI tools and applications that don't set a custom logging callback.Likely an incorrect or invalid review comment.
ggml/src/ggml-remotingfrontend/include/drm-uapi/drm.h (3)
66-1360: LGTM: DRM UAPI definitions are well-formed.The struct, enum, and constant definitions follow standard kernel UAPI conventions with appropriate types for kernel/user-space compatibility. However, ensure these definitions match the kernel version used by the remoting backend to avoid ABI mismatches.
1361-1408: LGTM: Typedef aliases and header closing are correct.The typedef aliases provide backward compatibility with older code, and the header closing with extern "C" and include guard is proper.
1025-1025: ****The
#include "drm_mode.h"at line 1025 is not a missing dependency issue. This is a standard Linux kernel DRM UAPI header expected to come from the system's libdrm development package. The CMakeLists.txt correctly configures this by:
- Adding
/usr/include/libdrm/to include directories- Linking against the system
drmlibrary- Documenting the dependency:
# dnf install -y libdrm-develThe repository intentionally omits standard kernel UAPI headers and only provides vendor-specific extensions (like
virtgpu_drm.h). When libdrm-devel is installed, the compiler will locatedrm_mode.hin the system include path during the local search initiated by the quoted include directive. This is proper design practice.Likely an incorrect or invalid review comment.
ggml/src/ggml-remotingfrontend/include/drm-uapi/virtgpu_drm.h (2)
40-227: LGTM: Virtgpu struct and constant definitions are well-formed.The virtgpu-specific command codes, flags, structs, and constants follow proper kernel UAPI conventions with appropriate sized types and clear field documentation.
229-276: LGTM: IOCTL definitions and header closing are correct.The IOCTL definitions properly use DRM_COMMAND_BASE from drm.h and follow standard kernel UAPI patterns. Header closing is well-formed.
OWNERS (1)
1-13: LGTM!The OWNERS file follows standard format and clearly defines approvers and reviewers.
ggml/src/ggml-remotingbackend/CMakeLists.txt (1)
1-21: LGTM!The CMake configuration correctly sets up the remoting backend library with appropriate C++20 compilation flags.
.gitignore (1)
47-47: LGTM!The simplified
build-*pattern is clearer and aligns well with the new build scripts introduced in this PR.ggml/src/ggml-metal/ggml-metal-device.m (1)
336-338: LGTM!Disabling verbose debug logging in the pipeline compilation path is appropriate for reducing log noise in production builds.
Also applies to: 361-365
ggml/src/ggml-remotingfrontend/virtgpu-forward-metal.cpp (1)
3-20: Verify error handling is sufficient.The function returns
trueunconditionally (line 19), even though theREMOTE_CALLmacro populates aretvariable (line 11) that appears to indicate success or failure. If the macro doesn'tFATALon all error conditions, the function might return success after a failed remote call.Based on the macro definition in
ggml/src/ggml-remotingfrontend/virtgpu-forward-impl.h, theREMOTE_CALLmacro does callFATALon errors (lines 25, 28), so the unconditional return should be safe. However, please verify that all error paths are covered by the macro's validation.ggml/src/ggml-remotingbackend/shared/venus_cs_ggml-rpc.h (1)
1-45: LGTM!The header properly declares the RPC serialization interface. The explicit padding field (line 21) is good practice for ensuring consistent struct layout across different compilers and platforms.
src/llama-kv-cache.cpp (1)
126-128: LGTM!Disabling the verbose per-layer device logging is consistent with the PR's goal to reduce noisy debug output. The logging can be re-enabled if needed by uncommenting these lines.
CMakePresets.json (1)
33-34: LGTM!The new remoting presets follow the established pattern and correctly enable the corresponding cache variables for the remoting frontend and backend builds.
ggml/CMakeLists.txt (1)
322-323: Public header exposure LGTM.Adding include/ggml-remoting-frontend.h to public headers looks correct.
Please confirm the header actually exists at ggml/include/ggml-remoting-frontend.h in this PR.
ggml/src/ggml-backend-reg.cpp (1)
76-79: Remoting frontend include/registration LGTM; verify define wiring.Use of GGML_USE_REMOTINGFRONTEND matches ggml_add_backend expectations, contingent on the CMake alias fix.
Also applies to: 207-210
ggml/src/ggml-remotingfrontend/ggml-metal-remoting.h (1)
1-16: LGTM!The header is well-structured with clear type definitions and function declarations for the Metal remoting backend interface.
|
🟢 Test of 'mac_ai test test_ci' succeeded after 00 hours 09 minutes 53 seconds. 🟢 • Link to the test results. • Link to the reports index. Test configuration: |
|
test passed ❤️ , merging |
1 participant
Summary by CodeRabbit
New Features
Chores