streamkit/Dockerfile.full-gpu at main · streamer45/streamkit · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
# SPDX-FileCopyrightText: © 2025 StreamKit Contributors
#
# SPDX-License-Identifier: MPL-2.0

# Multi-stage build for StreamKit server with GPU support
# syntax=docker/dockerfile:1

# Version configuration
# Update these versions to upgrade dependencies across all build stages
# Example: docker build --build-arg SHERPA_ONNX_VERSION=1.12.18 -f Dockerfile.gpu -t streamkit:gpu .
ARG SHERPA_ONNX_VERSION=1.12.17
# IMPORTANT: Use CUDA 12.3.2 with cuDNN 9 to match sherpa-onnx pre-built GPU binaries
# This is the recommended version for ONNX Runtime GPU in 2025
ARG CUDA_VERSION=12.3.2

# Stage 1: Build Rust dependencies
FROM rust:1.92-slim-bookworm AS rust-deps

WORKDIR /build

# Install build dependencies (includes all deps needed by workspace plugins)
RUN apt-get update && apt-get install -y \
    pkg-config \
    libssl-dev \
    g++ \
    cmake \
    libopus-dev \
    libvpx-dev \
    libclang-dev \
    clang \
    curl \
    git \
    && rm -rf /var/lib/apt/lists/*

# Copy workspace files to build dependencies (plugins are now excluded)
COPY Cargo.toml Cargo.lock ./
COPY apps ./apps
COPY crates ./crates
COPY sdks ./sdks
COPY wit ./wit
COPY assets ./assets

# Create dummy ui/dist directory so server's RustEmbed doesn't fail
# (will be replaced with real UI in Stage 3)
RUN mkdir -p ui/dist && echo '<!DOCTYPE html><html><body>Building...</body></html>' > ui/dist/index.html

# Build dependencies with cache mount (plugins are excluded from workspace)
RUN --mount=type=cache,target=/usr/local/cargo/registry \
    --mount=type=cache,target=/usr/local/cargo/git \
    --mount=type=cache,target=/build/target \
    cargo build --profile release-lto -p streamkit-server --bin skit --features "moq" -j$(nproc) && \
    # Copy compiled artifacts out of cache mount so they persist in the layer
    mkdir -p /build/target-out && \
    cp -r /build/target/release-lto /build/target-out/

# Stage 2: Build UI
FROM oven/bun:1.3.5-alpine AS ui-builder

WORKDIR /build/ui

# Install UI dependencies
COPY ui/package.json ui/bun.lock* ./
RUN --mount=type=cache,target=/root/.bun/install/cache \
    bun install --frozen-lockfile

# Copy UI source and build
COPY ui/ ./
RUN bun run build

# Stage 3: Build final server binary with UI embedded
FROM rust:1.92-slim-bookworm AS rust-builder

WORKDIR /build

# Install build dependencies (same as stage 1, includes all deps needed by workspace)
RUN apt-get update && apt-get install -y \
    pkg-config \
    libssl-dev \
    g++ \
    cmake \
    libopus-dev \
    libvpx-dev \
    libclang-dev \
    clang \
    curl \
    git \
    && rm -rf /var/lib/apt/lists/*

# Copy workspace files (plugins are excluded from workspace)
COPY Cargo.toml Cargo.lock ./
COPY apps ./apps
COPY crates ./crates
COPY sdks ./sdks
COPY wit ./wit
COPY assets ./assets

# Copy built UI from stage 2
COPY --from=ui-builder /build/ui/dist ./ui/dist

# Remove the server binary from the previous stage to force rebuild with new UI
# We'll use cache mount with artifacts from stage 1
RUN --mount=type=cache,target=/usr/local/cargo/registry \
    --mount=type=cache,target=/usr/local/cargo/git \
    --mount=type=cache,target=/build/target \
    --mount=type=bind,from=rust-deps,source=/build/target-out/release-lto,target=/build/target-init \
    bash -c '\
      # Copy pre-built dependencies if target is empty (first build) \
      if [ ! -d "/build/target/release-lto/deps" ]; then \
        echo "Initializing target from cache..."; \
        mkdir -p /build/target/release-lto; \
        cp -r /build/target-init/* /build/target/release-lto/ || true; \
      fi; \
      # Remove server binary to force rebuild with new UI \
      rm -rf /build/target/release-lto/skit \
        /build/target/release-lto/skit.d \
        /build/target/release-lto/deps/streamkit_server-* \
        /build/target/release-lto/.fingerprint/streamkit-server-*; \
      # Build only the server binary \
      cargo build --profile release-lto --features "moq" --bin skit -j$(nproc); \
      # Copy final binary out of cache mount \
      mkdir -p /build/bin && cp /build/target/release-lto/skit /build/bin/skit \
    '

# Stage 4: Build Whisper plugin with CUDA support
ARG CUDA_VERSION
FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 AS whisper-builder

WORKDIR /build

# Install Rust
RUN apt-get update && apt-get install -y \
    curl \
    pkg-config \
    libssl-dev \
    g++ \
    cmake \
    libclang-dev \
    clang \
    git \
    && rm -rf /var/lib/apt/lists/* \
    && curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y --default-toolchain 1.92.0

ENV PATH="/root/.cargo/bin:${PATH}"

# whisper.cpp/ggml defaults to enabling `GGML_NATIVE` (i.e. `-march=native`) when not cross-compiling.
# That can produce binaries that crash with SIGILL on older CPUs. Setting SOURCE_DATE_EPOCH disables
# `GGML_NATIVE_DEFAULT` upstream, making the build portable by default.
ENV SOURCE_DATE_EPOCH=1

# Extra defense-in-depth: ensure the toolchain doesn't auto-enable newer x86 features.
ENV CFLAGS="-O3 -pipe -fPIC -march=x86-64 -mtune=generic" \
    CXXFLAGS="-O3 -pipe -fPIC -march=x86-64 -mtune=generic"

# Copy only what's needed to build whisper plugin
# Note: Cargo.toml needed for workspace dependency resolution in core/
COPY Cargo.toml Cargo.lock ./
COPY crates/core ./crates/core
COPY sdks/plugin-sdk ./sdks/plugin-sdk
COPY plugins/native/whisper ./plugins/native/whisper

# Enable CUDA feature in whisper-rs by creating a patch Cargo.toml
RUN cd plugins/native/whisper && \
    sed -i 's/whisper-rs = "0.15"/whisper-rs = { version = "0.15", features = ["cuda"] }/' Cargo.toml

# Build whisper plugin with CUDA support
RUN --mount=type=cache,id=whisper-cargo-registry,target=/root/.cargo/registry \
    --mount=type=cache,id=whisper-cargo-git,target=/root/.cargo/git \
    --mount=type=cache,id=whisper-target-portable,target=/build/plugins/native/whisper/target-portable \
    cd plugins/native/whisper && \
    cargo build --release --target-dir target-portable -j$(nproc) && \
    mkdir -p /build/plugins/native && \
    cp target-portable/release/libwhisper.so /build/plugins/native/

# Download Whisper models (base.en, base multilingual, and tiny.en) and Silero VAD model
RUN mkdir -p /build/models && \
    curl -L -o /build/models/ggml-base.en-q5_1.bin \
      https://huggingface.co/streamkit/whisper-models/resolve/main/ggml-base.en-q5_1.bin && \
    curl -L -o /build/models/ggml-base-q5_1.bin \
      https://huggingface.co/streamkit/whisper-models/resolve/main/ggml-base-q5_1.bin && \
    curl -L -o /build/models/ggml-tiny.en-q5_1.bin \
      https://huggingface.co/streamkit/whisper-models/resolve/main/ggml-tiny.en-q5_1.bin && \
    curl -L -o /build/models/silero_vad.onnx \
      https://huggingface.co/streamkit/whisper-models/resolve/main/silero_vad.onnx

# Stage 5: Build Kokoro TTS plugin with GPU support
ARG CUDA_VERSION
FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 AS kokoro-builder

WORKDIR /build

# Install dependencies
RUN apt-get update && apt-get install -y \
    curl \
    wget \
    bzip2 \
    pkg-config \
    libssl-dev \
    libclang-dev \
    clang \
    && rm -rf /var/lib/apt/lists/* \
    && curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y --default-toolchain 1.92.0

ENV PATH="/root/.cargo/bin:${PATH}"

# Download and install pre-built GPU-enabled sherpa-onnx
# Note: The pre-built sherpa-onnx GPU binaries bundle ONNX Runtime, so we don't need to install it separately
# Using CUDA 12.x build which includes ONNX Runtime with CUDA 12 and cuDNN 9 support
ARG SHERPA_ONNX_VERSION
RUN cd /tmp && \
    wget https://github.com/k2-fsa/sherpa-onnx/releases/download/v${SHERPA_ONNX_VERSION}/sherpa-onnx-v${SHERPA_ONNX_VERSION}-cuda-12.x-cudnn-9.x-linux-x64-gpu.tar.bz2 && \
    tar xf sherpa-onnx-v${SHERPA_ONNX_VERSION}-cuda-12.x-cudnn-9.x-linux-x64-gpu.tar.bz2 && \
    cp -r sherpa-onnx-v${SHERPA_ONNX_VERSION}-cuda-12.x-cudnn-9.x-linux-x64-gpu/lib/* /usr/local/lib/ && \
    cp -r sherpa-onnx-v${SHERPA_ONNX_VERSION}-cuda-12.x-cudnn-9.x-linux-x64-gpu/include/* /usr/local/include/ && \
    ldconfig && \
    rm -rf /tmp/sherpa-onnx-v${SHERPA_ONNX_VERSION}-cuda-12.x-cudnn-9.x-linux-x64-gpu*

# Copy only what's needed to build kokoro plugin
# Note: Cargo.toml needed for workspace dependency resolution in core/
COPY Cargo.toml Cargo.lock ./
COPY crates/core ./crates/core
COPY sdks/plugin-sdk ./sdks/plugin-sdk
COPY plugins/native/kokoro ./plugins/native/kokoro

# Build kokoro plugin
RUN --mount=type=cache,id=kokoro-cargo-registry,target=/root/.cargo/registry \
    --mount=type=cache,id=kokoro-cargo-git,target=/root/.cargo/git \
    --mount=type=cache,id=kokoro-target,target=/build/plugins/native/kokoro/target \
    cd plugins/native/kokoro && \
    RUSTFLAGS="-L /usr/local/lib" cargo build --release --target-dir target -j$(nproc) && \
    mkdir -p /build/plugins/native && \
    cp target/release/libkokoro.so /build/plugins/native/

# Download Kokoro TTS models
RUN mkdir -p /build/models && \
    cd /build/models && \
    curl -L -o kokoro-multi-lang-v1_1.tar.bz2 \
      https://huggingface.co/streamkit/kokoro-models/resolve/main/kokoro-multi-lang-v1_1.tar.bz2 && \
    tar xf kokoro-multi-lang-v1_1.tar.bz2 && \
    rm kokoro-multi-lang-v1_1.tar.bz2

# Stage 6: Build Piper TTS plugin with GPU support
ARG CUDA_VERSION
FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 AS piper-builder

WORKDIR /build

# Install dependencies
RUN apt-get update && apt-get install -y \
    curl \
    wget \
    bzip2 \
    pkg-config \
    libssl-dev \
    libclang-dev \
    clang \
    && rm -rf /var/lib/apt/lists/* \
    && curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y --default-toolchain 1.92.0

ENV PATH="/root/.cargo/bin:${PATH}"

# Copy sherpa-onnx from kokoro-builder
COPY --from=kokoro-builder /usr/local/lib/*.so* /usr/local/lib/
COPY --from=kokoro-builder /usr/local/include/sherpa-onnx /usr/local/include/sherpa-onnx
RUN ldconfig

# Copy only what's needed to build piper plugin
# Note: Cargo.toml needed for workspace dependency resolution in core/
COPY Cargo.toml Cargo.lock ./
COPY crates/core ./crates/core
COPY sdks/plugin-sdk ./sdks/plugin-sdk
COPY plugins/native/piper ./plugins/native/piper

# Build piper plugin
RUN --mount=type=cache,id=piper-cargo-registry,target=/root/.cargo/registry \
    --mount=type=cache,id=piper-cargo-git,target=/root/.cargo/git \
    --mount=type=cache,id=piper-target,target=/build/plugins/native/piper/target \
    cd plugins/native/piper && \
    RUSTFLAGS="-L /usr/local/lib" cargo build --release --target-dir target -j$(nproc) && \
    mkdir -p /build/plugins/native && \
    cp target/release/libpiper.so /build/plugins/native/

# Download Piper TTS models
RUN mkdir -p /build/models && \
    cd /build/models && \
    # English model (sherpa-onnx format)
    curl -L -o vits-piper-en_US-libritts_r-medium.tar.bz2 \
      https://huggingface.co/streamkit/piper-models/resolve/main/vits-piper-en_US-libritts_r-medium.tar.bz2 && \
    tar xf vits-piper-en_US-libritts_r-medium.tar.bz2 && \
    rm vits-piper-en_US-libritts_r-medium.tar.bz2 && \
    cd vits-piper-en_US-libritts_r-medium && \
    if [ ! -f "model.onnx" ] && [ -f "en_US-libritts_r-medium.onnx" ]; then \
        ln -sf en_US-libritts_r-medium.onnx model.onnx; \
    fi && \
    cd /build/models && \
    # Mexican Spanish model (sherpa-onnx format with espeak-ng-data)
    curl -L -o vits-piper-es_MX-claude-high.tar.bz2 \
      https://huggingface.co/streamkit/piper-models/resolve/main/vits-piper-es_MX-claude-high.tar.bz2 && \
    tar xf vits-piper-es_MX-claude-high.tar.bz2 && \
    rm vits-piper-es_MX-claude-high.tar.bz2 && \
    cd vits-piper-es_MX-claude-high && \
    if [ ! -f "model.onnx" ] && [ -f "es_MX-claude-high.onnx" ]; then \
        ln -sf es_MX-claude-high.onnx model.onnx; \
    fi

# Stage 7: Build SenseVoice STT plugin with GPU support
ARG CUDA_VERSION
FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 AS sensevoice-builder

WORKDIR /build

# Install dependencies
RUN apt-get update && apt-get install -y \
    curl \
    wget \
    bzip2 \
    pkg-config \
    libssl-dev \
    libclang-dev \
    clang \
    && rm -rf /var/lib/apt/lists/* \
    && curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y --default-toolchain 1.92.0

ENV PATH="/root/.cargo/bin:${PATH}"

# Copy sherpa-onnx from kokoro-builder
COPY --from=kokoro-builder /usr/local/lib/*.so* /usr/local/lib/
COPY --from=kokoro-builder /usr/local/include/sherpa-onnx /usr/local/include/sherpa-onnx
RUN ldconfig

# Copy only what's needed to build sensevoice plugin
# Note: Cargo.toml needed for workspace dependency resolution in core/
COPY Cargo.toml Cargo.lock ./
COPY crates/core ./crates/core
COPY sdks/plugin-sdk ./sdks/plugin-sdk
COPY plugins/native/sensevoice ./plugins/native/sensevoice

# Build sensevoice plugin
RUN --mount=type=cache,id=sensevoice-cargo-registry,target=/root/.cargo/registry \
    --mount=type=cache,id=sensevoice-cargo-git,target=/root/.cargo/git \
    --mount=type=cache,id=sensevoice-target,target=/build/plugins/native/sensevoice/target \
    cd plugins/native/sensevoice && \
    RUSTFLAGS="-L /usr/local/lib" cargo build --release --target-dir target -j$(nproc) && \
    mkdir -p /build/plugins/native && \
    cp target/release/libsensevoice.so /build/plugins/native/

# Download SenseVoice models
RUN mkdir -p /build/models && \
    cd /build/models && \
    curl -L -o sherpa-onnx-sense-voice-zh-en-ja-ko-yue-int8-2025-09-09.tar.bz2 \
      https://huggingface.co/streamkit/sensevoice-models/resolve/main/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-int8-2025-09-09.tar.bz2 && \
    tar xf sherpa-onnx-sense-voice-zh-en-ja-ko-yue-int8-2025-09-09.tar.bz2 && \
    rm sherpa-onnx-sense-voice-zh-en-ja-ko-yue-int8-2025-09-09.tar.bz2

# Stage 8: Build VAD plugin with GPU support
ARG CUDA_VERSION
FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 AS vad-builder

WORKDIR /build

# Install dependencies
RUN apt-get update && apt-get install -y \
    curl \
    wget \
    bzip2 \
    pkg-config \
    libssl-dev \
    libclang-dev \
    clang \
    && rm -rf /var/lib/apt/lists/* \
    && curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y --default-toolchain 1.92.0

ENV PATH="/root/.cargo/bin:${PATH}"

# Copy sherpa-onnx from kokoro-builder
COPY --from=kokoro-builder /usr/local/lib/*.so* /usr/local/lib/
COPY --from=kokoro-builder /usr/local/include/sherpa-onnx /usr/local/include/sherpa-onnx
RUN ldconfig

# Copy only what's needed to build vad plugin
# Note: Cargo.toml needed for workspace dependency resolution in core/
COPY Cargo.toml Cargo.lock ./
COPY crates/core ./crates/core
COPY sdks/plugin-sdk ./sdks/plugin-sdk
COPY plugins/native/vad ./plugins/native/vad

# Build vad plugin
RUN --mount=type=cache,id=vad-cargo-registry,target=/root/.cargo/registry \
    --mount=type=cache,id=vad-cargo-git,target=/root/.cargo/git \
    --mount=type=cache,id=vad-target,target=/build/plugins/native/vad/target \
    cd plugins/native/vad && \
    RUSTFLAGS="-L /usr/local/lib" cargo build --release --target-dir target -j$(nproc) && \
    mkdir -p /build/plugins/native && \
    cp target/release/libvad.so /build/plugins/native/

# Download ten-vad model
RUN mkdir -p /build/models && \
    curl -L -o /build/models/ten-vad.onnx \
      https://huggingface.co/streamkit/vad-models/resolve/main/ten-vad.onnx

# Stage 9: Build Matcha TTS plugin with GPU support
ARG CUDA_VERSION
FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 AS matcha-builder

WORKDIR /build

# Install dependencies
RUN apt-get update && apt-get install -y \
    curl \
    wget \
    bzip2 \
    pkg-config \
    libssl-dev \
    libclang-dev \
    clang \
    && rm -rf /var/lib/apt/lists/* \
    && curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y --default-toolchain 1.92.0

ENV PATH="/root/.cargo/bin:${PATH}"

# Copy sherpa-onnx from kokoro-builder
COPY --from=kokoro-builder /usr/local/lib/*.so* /usr/local/lib/
COPY --from=kokoro-builder /usr/local/include/sherpa-onnx /usr/local/include/sherpa-onnx
RUN ldconfig

# Copy only what's needed to build matcha plugin
# Note: Cargo.toml needed for workspace dependency resolution in core/
COPY Cargo.toml Cargo.lock ./
COPY crates/core ./crates/core
COPY sdks/plugin-sdk ./sdks/plugin-sdk
COPY plugins/native/matcha ./plugins/native/matcha

# Build matcha plugin
RUN --mount=type=cache,id=matcha-cargo-registry,target=/root/.cargo/registry \
    --mount=type=cache,id=matcha-cargo-git,target=/root/.cargo/git \
    --mount=type=cache,id=matcha-target,target=/build/plugins/native/matcha/target \
    cd plugins/native/matcha && \
    RUSTFLAGS="-L /usr/local/lib" cargo build --release --target-dir target -j$(nproc) && \
    mkdir -p /build/plugins/native && \
    cp target/release/libmatcha.so /build/plugins/native/

# Download Matcha TTS models
RUN mkdir -p /build/models && \
    cd /build/models && \
    curl -L -o matcha-icefall-en_US-ljspeech.tar.bz2 \
      https://huggingface.co/streamkit/matcha-models/resolve/main/matcha-icefall-en_US-ljspeech.tar.bz2 && \
    tar xf matcha-icefall-en_US-ljspeech.tar.bz2 && \
    rm matcha-icefall-en_US-ljspeech.tar.bz2 && \
    cd matcha-icefall-en_US-ljspeech && \
    curl -L -o vocos-22khz-univ.onnx \
      https://huggingface.co/streamkit/matcha-models/resolve/main/matcha-icefall-en_US-ljspeech/vocos-22khz-univ.onnx

# Stage 10: Build NLLB Translation plugin with GPU support
# NOTE: This stage explicitly waits for whisper-builder to complete first
# to avoid running two heavy Rust+CUDA builds in parallel (disk space constraint on CI runners)
ARG CUDA_VERSION
FROM nvidia/cuda:${CUDA_VERSION}-cudnn9-devel-ubuntu22.04 AS nllb-builder

# Force serialization: wait for whisper-builder to complete before starting nllb-builder
# This reduces peak disk usage on CI runners with limited space
COPY --from=whisper-builder /build/plugins/native/libwhisper.so /tmp/.wait-for-whisper

WORKDIR /build

# Install dependencies
RUN apt-get update && apt-get install -y \
    curl \
    wget \
    pkg-config \
    libssl-dev \
    g++ \
    cmake \
    python3 \
    python3-pip \
    libclang-dev \
    clang \
    && rm -rf /var/lib/apt/lists/* \
    && curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y --default-toolchain 1.92.0

ENV PATH="/root/.cargo/bin:${PATH}"

# Compile and install CTranslate2 with CUDA support from source
RUN apt-get update && apt-get install -y git && \
    cd /tmp && \
    git clone --depth 1 --branch v4.6.1 --recursive https://github.com/OpenNMT/CTranslate2.git && \
    cd CTranslate2 && \
    mkdir build && cd build && \
    cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=/usr/local -DWITH_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES="75;80;86;89;90" -DWITH_MKL=OFF -DWITH_DNNL=OFF -DOPENMP_RUNTIME=COMP .. && \
    make -j$(nproc) && \
    make install && \
    ldconfig && \
    rm -rf /tmp/CTranslate2 && \
    apt-get remove -y git && apt-get autoremove -y && rm -rf /var/lib/apt/lists/*

# Copy only what's needed to build nllb plugin
# Note: Cargo.toml needed for workspace dependency resolution in core/
COPY Cargo.toml Cargo.lock ./
COPY crates/core ./crates/core
COPY sdks/plugin-sdk ./sdks/plugin-sdk
COPY plugins/native/nllb ./plugins/native/nllb

# Enable CUDA features in ct2rs by patching Cargo.toml
RUN cd plugins/native/nllb && \
    sed -i 's/ct2rs = "0.9"/ct2rs = { version = "0.9", features = ["cuda", "cudnn"] }/' Cargo.toml

# Build nllb plugin with CUDA support
RUN --mount=type=cache,id=nllb-cargo-registry,target=/root/.cargo/registry \
    --mount=type=cache,id=nllb-cargo-git,target=/root/.cargo/git \
    --mount=type=cache,id=nllb-target,target=/build/plugins/native/nllb/target \
    cd plugins/native/nllb && \
    CMAKE_BUILD_PARALLEL_LEVEL=$(nproc) RUSTFLAGS="-L /usr/local/lib" cargo build --release --target-dir target -j$(nproc) && \
    mkdir -p /build/plugins/native && \
    cp target/release/libnllb.so /build/plugins/native/

# Download NLLB pre-converted models from Hugging Face
# - 3.3B-float16: GPU-optimized model (~6.6GB)
# - 600M-INT8: CPU fallback model (~1.2GB)
RUN PIP_BREAK_SYSTEM_PACKAGES=1 pip3 install --no-cache-dir huggingface-hub && \
    mkdir -p /build/models && \
    cd /build/models && \
    python3 -c "from huggingface_hub import snapshot_download; snapshot_download('entai2965/nllb-200-3.3B-ctranslate2-float16', local_dir='nllb-200-3.3B-ct2-float16', local_dir_use_symlinks=False)" && \
    python3 -c "from huggingface_hub import snapshot_download; snapshot_download('entai2965/nllb-200-distilled-600M-ctranslate2', local_dir='nllb-200-distilled-600M-ct2-int8', local_dir_use_symlinks=False)"

# Stage 11: Build Helsinki Translation plugin (Candle) and download OPUS-MT models
ARG CUDA_VERSION
FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 AS helsinki-builder

WORKDIR /build

# Install dependencies (Rust + Python for model conversion)
RUN apt-get update && apt-get install -y \
    curl \
    git \
    pkg-config \
    libssl-dev \
    libclang-dev \
    clang \
    python3 \
    python3-pip \
    && rm -rf /var/lib/apt/lists/* \
    && curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y --default-toolchain 1.92.0

ENV PATH="/root/.cargo/bin:${PATH}"

# bindgen_cuda (used by candle-kernels) tries to call `nvidia-smi` to infer the
# GPU compute capability. During `docker build`, GPUs are typically not exposed
# and `nvidia-smi` is not present, so we provide a default compute capability.
# Override at build time if desired, e.g.:
#   docker build --build-arg HELSINKI_CUDA_COMPUTE_CAP=86 ...
ARG HELSINKI_CUDA_COMPUTE_CAP=70
ENV CUDA_COMPUTE_CAP=${HELSINKI_CUDA_COMPUTE_CAP}

# Copy only what's needed to build helsinki plugin
COPY Cargo.toml Cargo.lock ./
COPY crates/core ./crates/core
COPY sdks/plugin-sdk ./sdks/plugin-sdk
COPY plugins/native/helsinki ./plugins/native/helsinki

# Build helsinki plugin
RUN --mount=type=cache,id=helsinki-cargo-registry,target=/root/.cargo/registry \
    --mount=type=cache,id=helsinki-cargo-git,target=/root/.cargo/git \
    --mount=type=cache,id=helsinki-target,target=/build/plugins/native/helsinki/target \
    cd plugins/native/helsinki && \
    cargo build --release --features cuda --target-dir target -j$(nproc) && \
    mkdir -p /build/plugins/native && \
    cp target/release/libhelsinki.so /build/plugins/native/

# Download and convert OPUS-MT models to Candle format (safetensors + tokenizer json)
RUN PIP_BREAK_SYSTEM_PACKAGES=1 pip3 install --no-cache-dir \
      transformers \
      sentencepiece \
      safetensors \
      torch \
      tokenizers && \
    python3 plugins/native/helsinki/download-models.py

# Runtime stage with NVIDIA CUDA runtime + cuDNN
# IMPORTANT: Use cudnn9-runtime for CUDA 12.x with ONNX Runtime CUDA support
ARG CUDA_VERSION
FROM nvidia/cuda:${CUDA_VERSION}-cudnn9-runtime-ubuntu22.04

# Install runtime dependencies and debugging tools
RUN apt-get update && apt-get install -y \
    ca-certificates \
    libssl3 \
    libopus0 \
    libvpx7 \
    libgomp1 \
    gdb \
    curl \
    iproute2 \
    && rm -rf /var/lib/apt/lists/*

# Create app user
RUN useradd -m -u 1000 -s /bin/bash app

# Copy binary from rust builder
COPY --from=rust-builder /build/bin/skit /usr/local/bin/skit

# Copy ONNX Runtime GPU libraries from kokoro-builder
COPY --from=kokoro-builder /usr/local/lib/*.so* /usr/local/lib/

# Copy CUDA libraries needed at runtime (cuDNN, cuBLAS, etc.)
# Note: libcuda.so.1 is provided by the host NVIDIA driver via nvidia-container-toolkit
# The runtime image includes other CUDA libraries (cuBLAS, cuDNN, etc.)
# Ensure libraries are in the linker cache
RUN ldconfig

# Set library path to include NVIDIA libraries
# The nvidia-container-toolkit will inject /usr/local/nvidia/lib64 at runtime
ENV LD_LIBRARY_PATH=/usr/local/lib:/usr/local/nvidia/lib64:/usr/local/nvidia/lib:${LD_LIBRARY_PATH}

# Copy whisper plugin and models
# IMPORTANT: Use --chown=app:app on every COPY to avoid a bulk `chown -R` later
# which would duplicate every file in a new layer (~12GB waste).
COPY --chown=app:app --from=whisper-builder /build/plugins /opt/streamkit/plugins
COPY --chown=app:app --from=whisper-builder /build/models /opt/streamkit/models

# Copy kokoro plugin and models (merge into /opt/streamkit/plugins and /opt/streamkit/models)
COPY --chown=app:app --from=kokoro-builder /build/plugins/native/* /opt/streamkit/plugins/native/
COPY --chown=app:app --from=kokoro-builder /build/models/kokoro-multi-lang-v1_1 /opt/streamkit/models/kokoro-multi-lang-v1_1

# Copy piper plugin and models (merge into /opt/streamkit/plugins and /opt/streamkit/models)
COPY --chown=app:app --from=piper-builder /build/plugins/native/* /opt/streamkit/plugins/native/
COPY --chown=app:app --from=piper-builder /build/models/vits-piper-en_US-libritts_r-medium /opt/streamkit/models/vits-piper-en_US-libritts_r-medium
COPY --chown=app:app --from=piper-builder /build/models/vits-piper-es_MX-claude-high /opt/streamkit/models/vits-piper-es_MX-claude-high

# Copy sensevoice plugin and models (merge into /opt/streamkit/plugins and /opt/streamkit/models)
COPY --chown=app:app --from=sensevoice-builder /build/plugins/native/* /opt/streamkit/plugins/native/
COPY --chown=app:app --from=sensevoice-builder /build/models/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-int8-2025-09-09 /opt/streamkit/models/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-int8-2025-09-09

# Copy CTranslate2 libraries from nllb-builder
COPY --from=nllb-builder /usr/local/lib/libctranslate2.so* /usr/local/lib/
RUN ldconfig

# Copy nllb plugin and models (merge into /opt/streamkit/plugins and /opt/streamkit/models)
# - 3.3B-float16: GPU-optimized model
# - 600M-INT8: CPU fallback model
COPY --chown=app:app --from=nllb-builder /build/plugins/native/* /opt/streamkit/plugins/native/
COPY --chown=app:app --from=nllb-builder /build/models/nllb-200-3.3B-ct2-float16 /opt/streamkit/models/nllb-200-3.3B-ct2-float16
COPY --chown=app:app --from=nllb-builder /build/models/nllb-200-distilled-600M-ct2-int8 /opt/streamkit/models/nllb-200-distilled-600M-ct2-int8

# Copy helsinki plugin and models (merge into /opt/streamkit/plugins and /opt/streamkit/models)
COPY --chown=app:app --from=helsinki-builder /build/plugins/native/* /opt/streamkit/plugins/native/
COPY --chown=app:app --from=helsinki-builder /build/models/opus-mt-en-es /opt/streamkit/models/opus-mt-en-es
COPY --chown=app:app --from=helsinki-builder /build/models/opus-mt-es-en /opt/streamkit/models/opus-mt-es-en

# Copy vad plugin and model (merge into /opt/streamkit/plugins and /opt/streamkit/models)
COPY --chown=app:app --from=vad-builder /build/plugins/native/* /opt/streamkit/plugins/native/
COPY --chown=app:app --from=vad-builder /build/models/ten-vad.onnx /opt/streamkit/models/ten-vad.onnx

# Copy matcha plugin and models (merge into /opt/streamkit/plugins and /opt/streamkit/models)
COPY --chown=app:app --from=matcha-builder /build/plugins/native/* /opt/streamkit/plugins/native/
COPY --chown=app:app --from=matcha-builder /build/models/matcha-icefall-en_US-ljspeech /opt/streamkit/models/matcha-icefall-en_US-ljspeech

# Copy sample pipelines
COPY --chown=app:app samples /opt/streamkit/samples

# Copy official Docker configuration (no dependency on private demo assets)
COPY --chown=app:app docker-skit.toml /opt/streamkit/config/skit.toml
COPY --chown=app:app docker-skit-gpu.toml /opt/streamkit/config/skit-gpu.toml

# Copy demo configurations used by `demo/launcher`
COPY --chown=app:app demo/skit.toml /opt/streamkit/demo/skit.toml
COPY --chown=app:app demo/skit-gpu.toml /opt/streamkit/demo/skit-gpu.toml

# Create runtime-writable directories and compatibility symlinks.
# No bulk chown needed — all COPY instructions above use --chown=app:app.
RUN mkdir -p /opt/streamkit/.plugins /opt/streamkit/logs /opt/streamkit/config /opt/streamkit/demo && \
    chown app:app /opt/streamkit/.plugins /opt/streamkit/logs /opt/streamkit/config /opt/streamkit/demo && \
    mkdir -p /app && \
    ln -sfn /opt/streamkit/config/skit.toml /opt/streamkit/skit.toml && \
    ln -sfn /opt/streamkit/config/skit-gpu.toml /opt/streamkit/skit-gpu.toml && \
    ln -sfn /opt/streamkit/plugins /app/plugins && \
    ln -sfn /opt/streamkit/models /app/models && \
    ln -sfn /opt/streamkit/samples /app/samples && \
    ln -sfn /opt/streamkit/demo /app/demo && \
    ln -sfn /opt/streamkit/config /app/config && \
    ln -sfn /opt/streamkit/.plugins /app/.plugins && \
    ln -sfn /opt/streamkit/logs /app/logs

WORKDIR /opt/streamkit
USER app

# Expose HTTP and UDP ports
EXPOSE 4545/tcp
EXPOSE 4545/udp

# OCI image labels
LABEL org.opencontainers.image.title="StreamKit (GPU)"
LABEL org.opencontainers.image.description="High-performance real-time media processing engine with GPU-accelerated plugins (CUDA 12.3)"
LABEL org.opencontainers.image.source="https://github.com/streamer45/streamkit"
LABEL org.opencontainers.image.licenses="MPL-2.0"
LABEL org.opencontainers.image.vendor="StreamKit Contributors"

# Health check
HEALTHCHECK --interval=30s --timeout=3s --start-period=10s --retries=3 \
    CMD curl -f http://localhost:4545/healthz || exit 1

# Default command
CMD ["skit", "serve"]