[ExecuTorch][WebGPU] Add Gemma 4 plain export and artifact contract

backends/webgpu/CMakeLists.txt

@@ -1,3 +1,3 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
 #
@@ -116,6 +116,51 @@
 
 set_property(TARGET webgpu_backend PROPERTY CXX_STANDARD 17)
 
+if(EMSCRIPTEN)
+  add_executable(
+    gemma4_plain_wasm
+    ${EXECUTORCH_ROOT}/examples/models/gemma4/runner/gemma4_plain_wasm.cpp
+  )
+  target_include_directories(
+    gemma4_plain_wasm PRIVATE $<BUILD_INTERFACE:${EXECUTORCH_ROOT}/..>
+  )
+  target_link_libraries(
+    gemma4_plain_wasm PRIVATE webgpu_backend webgpu_model_loader extension_tensor
+  )
+  target_compile_options(
+    gemma4_plain_wasm PRIVATE -fexceptions "--use-port=emdawnwebgpu"
+  )
+  if(EXECUTORCH_BUILD_WEBGPU_PROFILING)
+    target_compile_definitions(
+      gemma4_plain_wasm PRIVATE WGPU_BACKEND_ENABLE_PROFILING
+    )
+  endif()
+  target_link_options(
+    gemma4_plain_wasm
+    PRIVATE
+    -fexceptions
+    "--use-port=emdawnwebgpu"
+    -sASYNCIFY
+    -sALLOW_MEMORY_GROWTH=1
+    -sMAXIMUM_MEMORY=4GB
+    -sFORCE_FILESYSTEM=1
+    --no-entry
+    "-sEXPORTED_FUNCTIONS=_et_init,_et_load,_et_unload,_et_reset,_et_prefill_batch,_et_prefill_step,_et_step,_et_profile_enable,_et_profile,_et_get_last_prefill_token_count,_et_get_route_contract_version,_et_get_last_route_mask,_et_get_last_route_conflict_count,_malloc,_free"
+    "-sEXPORTED_RUNTIME_METHODS=ccall,cwrap,FS,HEAP32"
+    -sSTACK_SIZE=8388608
+    -sASYNCIFY_STACK_SIZE=1048576
+    -sMODULARIZE=1
+    -sEXPORT_NAME=createWebGPULlama
+  )
+  set_target_properties(
+    gemma4_plain_wasm
+    PROPERTIES OUTPUT_NAME "webgpu_llama"
+               RUNTIME_OUTPUT_DIRECTORY
+               "${CMAKE_CURRENT_BINARY_DIR}/browser_gemma4_plain"
+               CXX_STANDARD 17
+  )
+endif()
+
 install(
   TARGETS webgpu_backend
   EXPORT ExecuTorchTargets

examples/models/gemma4/BUCK

@@ -1,13 +1,15 @@
 load("@fbcode_macros//build_defs:build_file_migration.bzl", "fbcode_target", "non_fbcode_target")
 load("@fbsource//xplat/executorch/build:runtime_wrapper.bzl", "runtime")
-load(":targets.bzl", "define_common_targets")
+load(":targets.bzl", "define_common_targets", "define_webgpu_python_targets")
 
 oncall("executorch")
 
 non_fbcode_target(_kind = define_common_targets,)
 
 fbcode_target(_kind = define_common_targets,)
 
+define_webgpu_python_targets()
+
 # Text decoder module
 fbcode_target(_kind = runtime.python_library,
     name = "text_decoder",
@@ -77,6 +79,7 @@ fbcode_target(_kind = runtime.python_binary,
         "//executorch/kernels/quantized:aot_lib",
     ],
     deps = [
+        ":webgpu_support",
         ":text_decoder",
         ":speech_transform",
         ":quant_utils",

examples/models/gemma4/README.md

@@ -45,6 +45,84 @@ buck2 run fbcode//executorch/examples/models/gemma4:export_gemma4 -- \
     --checkpoint_path /tmp/gemma4-e2b-it --no-audio
 ```
 
+### Plain E2B WebGPU
+
+The WebGPU path is independently exportable and text-only. It preserves an
+8960-token KV capacity while bounding each input call to 512 tokens, returns a
+delegated `Long[1, 1]` greedy token, and splits external constants into three
+ordered PTD files below the browser binding/fetch limit. The default XNNPACK
+export and runner are unchanged.
+
+Acquire the checkpoint from
+`google/gemma-4-E2B-it-qat-q4_0-unquantized` at immutable revision
+`6befbaca7398925921802abd1f277b495b78b738`, then validate every staged byte:
+
+```bash
+hf download google/gemma-4-E2B-it-qat-q4_0-unquantized \
+    model.safetensors config.json tokenizer.json tokenizer_config.json \
+    generation_config.json processor_config.json chat_template.jinja \
+    README.md .gitattributes \
+    --revision 6befbaca7398925921802abd1f277b495b78b738 \
+    --local-dir /tmp/gemma4-e2b-it
+```
+
+```bash
+buck2 run fbcode//executorch/examples/models/gemma4:webgpu_artifact_manifest -- \
+    validate-acquisition --checkpoint-root /tmp/gemma4-e2b-it
+```
+
+From clean fbsource and ExecuTorch OSS checkouts, seal the reviewed plain-Gemma
+source union and generator-derived WGSL closure before exporting the model:
+
+```bash
+: "${FBSOURCE_ROOT:?set the clean fbsource checkout root}"
+: "${OSS_ROOT:?set the clean ExecuTorch OSS checkout root}"
+python -m executorch.examples.models.gemma4.webgpu_artifact_manifest \
+  create-source-manifest --fbsource-root "$FBSOURCE_ROOT" \
+  --oss-root "$OSS_ROOT" \
+  --output /tmp/gemma4-source-manifest.json
+python -m executorch.examples.models.gemma4.webgpu_artifact_manifest \
+  create-wgsl-manifest \
+  --backend-root "$FBSOURCE_ROOT/xplat/executorch/backends/webgpu" \
+  --output /tmp/gemma4-wgsl-manifest.json
+python -m executorch.examples.models.gemma4.webgpu_artifact_manifest \
+  create-source-receipt \
+  --fbsource-root "$FBSOURCE_ROOT" --oss-root "$OSS_ROOT" \
+  --backend-root "$FBSOURCE_ROOT/xplat/executorch/backends/webgpu" \
+  --output /tmp/gemma4-source-receipt.json
+```
+
+Export plain Gemma 4:
+
+```bash
+buck2 run fbcode//executorch/examples/models/gemma4:export_gemma4 -- \
+    --checkpoint_path /tmp/gemma4-e2b-it \
+    --output_path /tmp/gemma4-webgpu/model.pte \
+    --backend webgpu --quantize 8da4w+emb4 \
+    --max_seq_len 8960 --max_input_len 512 \
+    --no-audio --no-vision \
+    --source_receipt_path /tmp/gemma4-source-receipt.json \
+    --artifact_manifest_output /tmp/gemma4-e2b-webgpu.json
+```
+
+The source-closure gate writes `gemma4-source-receipt.json` before this export.
+The exporter reads the actual tensor-data insertion order, writes all three
+content-named PTDs, and creates the manifest without renaming or globbing them.
+Keep the manifest output outside the flat artifact staging directory.
+
+```bash
+buck2 run fbcode//executorch/examples/models/gemma4:webgpu_artifact_manifest -- \
+    validate --root /tmp/gemma4-webgpu \
+    --manifest /tmp/gemma4-e2b-webgpu.json
+```
+
+`manifests/gemma4_e2b_webgpu.json` pins the accepted ctx8960 behavior-oracle
+quartet, but marks its old-worktree source closure as pending. It cannot satisfy
+the production validator because it has no final-source receipt. Rebuild it
+from the reviewed stack before claiming source-current performance or
+reproduction. Dashboard and internal publication paths are evidence only and
+are never source dependencies.
+
 ## Model Variants
 
 Default export includes all modalities (audio + vision + text). Default context length: 1024 tokens (`--max_seq_len`).