nodes/hao-ai-lab-FastVideo
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[misc] Move environment var reads into env registry and rename unprefixed variables (#1897)

Browse Source
This commit is contained in:
Junda Su
2026-09-29 14:36:16 -07:00
committed by GitHub
parent 7f03e03dc6
commit 02f1ce11ae
54 changed files with 660 additions and 485 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.agents/skills/env-var-conventions/SKILL.md
+5
View File
@@ -34,10 +34,15 @@ on a summary here.
1. **Declare or edit the variable in `fastvideo/envs.py`.**
- Pick the field type and category that the policy doc lists.
- Write a description that states what the variable does and its units.
- To rename, keep the old name in `deprecated_names`. To remove, add the
name to `DEPRECATED_VARIABLES`. Update the uses in `examples/`,
`scripts/`, `docs/`, `apps/`, and the tests.
2. **Read the variable with `envs.NAME.get()` inside a function.**
- In tests, change the value with `envs.NAME.override(value)`.
- Do not call `os.environ`, `os.getenv`, or `monkeypatch.setenv` for a
FastVideo variable.
- To set a variable that another tool reads, call `envs.set_external`,
`envs.setdefault_external`, or `envs.unset_external`.
3. **Regenerate the table in the policy doc.**
- Run `python fastvideo/tests/contract/test_env_policy.py`.
4. **Run the contract test.**
docs/contributing/activation_trace.md
+1 -1
View File
@@ -30,7 +30,7 @@ FastVideo and the reference model first produce different numbers?"
| General logging | `init_logger(__name__)` |
| Per-stage timing | `FASTVIDEO_STAGE_LOGGING` |
| Profiling kernel timings | `FASTVIDEO_TORCH_PROFILER_DIR` (see [Profiling](profiling.md)) |
| Function-call tracing | `FASTVIDEO_TRACE_FUNCTION` (heavy) |
| Function-call tracing | `fastvideo.logger.enable_trace_function_call()` (heavy) |
## Quickstart
docs/contributing/env_vars.md
+123 -61
View File
@@ -11,14 +11,16 @@ test in the same pull request.
1. **Register every FastVideo variable in `fastvideo/envs.py`.** Each entry declares a type, a default, a category,
and a description. Variables that other tools own (CUDA, NCCL, PyTorch, launchers) are not registered; code reads
them directly with `os.environ.get("NAME")`, and the name must be in the external-variable allowlist
(`EXTERNAL_ALLOWLIST` in the contract test).
(`EXTERNAL_ALLOWLIST` in the contract test). When FastVideo sets such a variable for the other tool, it calls
`envs.set_external`, `envs.setdefault_external`, or `envs.unset_external`, and the name must be in
`EXTERNAL_WRITE_ALLOWLIST`.
2. **Read with `envs.NAME.get()`, write with `envs.NAME.set()`, and change a value in tests with
`envs.NAME.override()`.** Each type has one parsing rule. A value that the rule rejects raises
`fastvideo.envs.EnvVarError` instead of falling back to the default.
3. **Name FastVideo variables with the `FASTVIDEO_` prefix.** The second word states the purpose where one applies:
`ENABLE_`, `DISABLE_`, `USE_`, `FORCE_`, `DEBUG_`, `TEST_`.
4. **Keep a renamed variable as a deprecated alias until the next minor release.** Setting the old name logs a
warning. Delete a variable that no code reads, and list it in a deprecation table so that setting it logs a
warning. Delete a variable that no code reads, and list it in `DEPRECATED_VARIABLES` so that setting it logs a
warning.
5. **Give each setting one source: an argument or an environment variable.** Settings that users change per
deployment are arguments (CLI or YAML). Expert switches, emergency off switches, and debugging and test switches
@@ -85,6 +87,25 @@ with envs.FASTVIDEO_DEBUG_MY_STAGE.override(True):
run_stage()
```
## Rename or remove a variable
To rename a variable, declare it under the new name and list the old name in `deprecated_names`:
```python
FASTVIDEO_LTX2_USE_DISTILLED_SIGMAS = EnvBool(True,
category="sampling",
doc="...",
deprecated_names=("LTX2_USE_DISTILLED_SIGMAS", ))
```
`get()` reads an old name only when the new name is unset, and logs a warning once. Update the uses of the old name
in `examples/`, `scripts/`, `docs/`, `apps/`, and the tests in the same pull request. Delete the old name in the next
minor release.
To remove a variable that no code reads, delete its entry and add the name to `DEPRECATED_VARIABLES` in
`fastvideo/envs.py` with a reason. `FastVideoArgs` calls `envs.warn_deprecated_variables()`, which logs a warning for
each listed variable that is set. Delete the entry in the next minor release.
## What the contract test checks
The test parses every Python file under `fastvideo/`, including `fastvideo/tests/`, with Python's `ast` module. It
@@ -100,20 +121,24 @@ It reports each violation as `<path>: <kind> <name>`:
| | with a name built at runtime (`<dynamic>`) | another tool owns, add it to |
| | | `EXTERNAL_ALLOWLIST` with a reason. |
| `write` | `os.environ[...] = ...`, `setdefault`, `pop`, `del`, | Pass an argument instead. In tests, use |
| | `os.putenv`, `os.unsetenv`, `monkeypatch.setenv`/`delenv` | `envs.NAME.override()`. |
| | `os.putenv`, `os.unsetenv`, `monkeypatch.setenv`/`delenv`, | `envs.NAME.override()`. For a variable that |
| | or an `envs.*_external` call with a name outside | another tool reads, call an |
| | `EXTERNAL_WRITE_ALLOWLIST` | `envs.*_external` helper and add the name to |
| | | `EXTERNAL_WRITE_ALLOWLIST` with a reason. |
| `whole-environ` | `os.environ.copy()`, `dict(os.environ)`, iteration, | Read the specific variables that the code |
| | `mock.patch.dict(os.environ, ...)`, `os.environ.update` | needs. |
| `bare-field` | A registry field used without calling one of its methods, | Call `envs.NAME.get()`. |
| | as in `envs.NAME == "auto"` or `getter = envs.NAME.get` | |
| `import-time-read` | `envs.NAME.get()` outside a function | Move the read into the function that uses |
| | | the value. |
| `prefix` | A registry entry without the `FASTVIDEO_` prefix | Rename the variable and keep the old name as |
| | | a deprecated alias. |
| `unread` | A registry entry that no code reads with `get()` or | Delete the variable and add it to the |
| | `is_set()` | deprecation table. |
| `prefix` | A registry entry without the `FASTVIDEO_` prefix | Rename the variable and keep the old name in |
| | | `deprecated_names`. |
| `unread` | A registry entry that no code reads with `get()` or | Delete the variable and add it to |
| | `is_set()` | `DEPRECATED_VARIABLES`. |
The test recognizes `os` imported under another name and `from os import environ, getenv`. Code that reaches the
environment through `importlib` or `getattr(os, "environ")` is left to code review.
The test recognizes `os` imported under another name, `from os import environ, getenv`, and a name held in a
module-level string constant. Code that reaches the environment through `importlib` or `getattr(os, "environ")` is
left to code review.
The test also checks that every registry entry has a category from `envs.CATEGORIES` and a description, and that the
table at the end of this page matches the registry.
@@ -125,56 +150,93 @@ longer exists also fails the test, so the fixing pull request deletes its entry.
## Registered variables
<!-- BEGIN GENERATED ENV TABLE: python fastvideo/tests/contract/test_env_policy.py -->
| Variable | Type | Default | Category | Description |
| ---------------------------------------------- | ------------------------------ | --------------------------- | ----------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `FASTVIDEO_TARGET_DEVICE` | str | `cuda` | build | Target device of FastVideo: cuda, rocm, neuron, cpu, or openvino. |
| `MAX_JOBS` | str | unset | build | Maximum number of parallel compilation jobs. Defaults to the number of CPUs. |
| `NVCC_THREADS` | str | unset | build | Number of nvcc threads. When set, MAX_JOBS is reduced to avoid oversubscribing the CPU. |
| `FASTVIDEO_USE_PRECOMPILED` | bool | `0` | build | Use precompiled binaries (\*.so). |
| `CMAKE_BUILD_TYPE` | str | unset | build | CMake build type: Debug, Release, or RelWithDebInfo. |
| `VERBOSE` | bool | `0` | build | Print verbose logs during installation. |
| `FASTVIDEO_CONFIG_ROOT` | path | computed | path | Root directory for FastVideo configuration files, at runtime and at installation. Defaults to ~/.config/fastvideo, or $XDG_CONFIG_HOME/fastvideo when XDG_CONFIG_HOME is set. |
| `FASTVIDEO_CACHE_ROOT` | path | computed | path | Root directory for FastVideo cache files. Defaults to ~/.cache/fastvideo, or $XDG_CACHE_HOME/fastvideo when XDG_CACHE_HOME is set. |
| `FASTVIDEO_HOST_IP` | str | `""` | distributed | IP address of this node when the node has several network interfaces. Set it on each node for multi-node inference. |
| `FASTVIDEO_LOOPBACK_IP` | str | `""` | distributed | Loopback IP address to use instead of the detected one. |
| `FASTVIDEO_RAY_PER_WORKER_GPUS` | float | `1.0` | distributed | GPUs per Ray worker. A fraction lets Ray schedule several actors on one GPU, so other actors can share the GPUs with FastVideo. |
| `FASTVIDEO_RINGBUFFER_WARNING_INTERVAL` | int | `60` | distributed | Seconds between warnings while the ring buffer is full. |
| `FASTVIDEO_NCCL_SO_PATH` | str | unset | distributed | Path to the NCCL library file. Needed because the nccl&gt;=2.19 that PyTorch ships has a bug (https://github.com/NVIDIA/nccl/issues/1234). |
| `HCCL_SO_PATH` | str | unset | distributed | Path to the HCCL library file on Ascend NPUs. |
| `FASTVIDEO_ENGINE_ITERATION_TIMEOUT_S` | int | `60` | distributed | Timeout in seconds for each engine iteration. |
| `FASTVIDEO_WORKER_MULTIPROC_METHOD` | one of spawn, fork, forkserver | `spawn` | distributed | Multiprocessing start method for worker processes. |
| `FASTVIDEO_ULYSSES_A2A` | one of off, auto | `off` | distributed | Sequence-parallel all-to-all backend. off uses the NCCL path in DistributedAutograd.AllToAll4D. auto uses the fused NVLink kernel when the group is a load-store accessible mesh of 2, 4, 6, or 8 ranks in eager execution, and the NCCL path otherwise. |
| `LD_LIBRARY_PATH` | str | unset | external | Searched for the NCCL library when FASTVIDEO_NCCL_SO_PATH is unset. |
| `LOCAL_RANK` | int | `0` | external | Local rank of the process in a distributed run; selects the GPU device id. |
| `CUDA_VISIBLE_DEVICES` | str | unset | external | Visible devices in a distributed run. |
| `FASTVIDEO_CONFIGURE_LOGGING` | bool | `1` | logging | Configure logging at import. When true, FastVideo uses its default logging configuration or the file in FASTVIDEO_LOGGING_CONFIG_PATH. |
| `FASTVIDEO_LOGGING_CONFIG_PATH` | str | unset | logging | Path to a JSON logging configuration file. |
| `FASTVIDEO_LOGGING_LEVEL` | str | `INFO` | logging | Default logging level. |
| `FASTVIDEO_LOGGING_PREFIX` | str | `""` | logging | Prefix prepended to every log message. |
| `FASTVIDEO_STAGE_LOGGING` | bool | `0` | logging | Log the time that each pipeline stage takes. |
| `FASTVIDEO_ATTENTION_BACKEND` | str | unset | attention | Attention backend, as an AttentionBackendEnum name such as TORCH_SDPA, FLASH_ATTN, VIDEO_SPARSE_ATTN, SAGE_ATTN, or SAGE_ATTN_THREE. FastVideoArgs uses it when FastVideoArgs.attention_backend is unset. |
| `FASTVIDEO_FA4` | bool | `0` | attention | The FLASH_ATTN backend uses FlashAttention-4 (flash_attn.cute) instead of FA3 or FA2. |
| `FASTVIDEO_MINIMAX_H3_FA4_PACKED_VARLEN` | bool | `0` | attention | MiniMax-H3 dense DiT self-attention uses the FlashAttention-4 packed-varlen entry point. This changes the floating-point reduction order, so it is an inference-only opt-in. |
| `FASTVIDEO_VSA_SM100A` | bool | `0` | attention | VIDEO_SPARSE_ATTN_H3 sends no-grad tile-64 forwards to the data-center Blackwell (sm_100a) kernel. fastvideo-kernel reads the same variable with the same rule. |
| `FASTVIDEO_INFERENCE_TORCH_COMPILE` | bool | `0` | performance | Compile each DiT transformer block with fullgraph torch.compile at inference. Same as FastVideoArgs.inference_torch_compile=True. |
| `FASTVIDEO_VAE_PARALLEL_DECODE` | bool | `0` | performance | MiniMax-H3 VAE decode splits its temporal chunks across the sequence-parallel ranks instead of running serially on the output rank. Same as FastVideoArgs.vae_parallel_decode=True. |
| `FASTVIDEO_VAE_PARALLEL_ENCODE` | bool | `0` | performance | MiniMax-H3 reference-video VAE encode splits its temporal chunks across the sequence-parallel ranks. Same as FastVideoArgs.vae_parallel_encode=True. |
| `FASTVIDEO_VAE_PARALLEL_DECODE_STRATEGY` | str | unset | performance | Collective that moves chunks in parallel VAE decode: gather (used when unset) or all_gather. |
| `FASTVIDEO_MINIMAX_H3_FUSIONS` | str | `""` | performance | MiniMax-H3 inference-only Triton fusions: all, 1, or a comma-separated subset of modulate,qknorm_rope,swiglu. Empty, 0, or none keeps the eager implementation. |
| `FASTVIDEO_TEST_DYNAMO_FULLGRAPH_CAPTURE` | bool | `1` | debug | Enable Dynamo fullgraph capture. |
| `FASTVIDEO_NVTX_PROFILE` | bool | `0` | profiling | Emit NVTX ranges for external profilers such as Nsight Systems. |
| `FASTVIDEO_TORCH_PROFILER_DIR` | path | unset | profiling | Enables the torch profiler and sets the directory for its traces. Must be an absolute path. |
| `FASTVIDEO_TORCH_PROFILER_RECORD_SHAPES` | bool | `0` | profiling | Torch profiler records shapes. |
| `FASTVIDEO_TORCH_PROFILER_WITH_PROFILE_MEMORY` | bool | `0` | profiling | Torch profiler profiles memory. |
| `FASTVIDEO_TORCH_PROFILER_WITH_STACK` | bool | `0` | profiling | Torch profiler captures stacks. Costs about 1.5x runtime and 1.4x trace size. |
| `FASTVIDEO_TORCH_PROFILER_WITH_FLOPS` | bool | `0` | profiling | Torch profiler profiles FLOPs. |
| `FASTVIDEO_TORCH_PROFILE_REGIONS` | str | `""` | profiling | Comma-separated profiler regions to record. The torch profiler requires at least one region. |
| `FASTVIDEO_SERVER_DEV_MODE` | bool | `0` | debug | Run the server in development mode with extra debugging endpoints. |
| `FASTVIDEO_TRACE_FUNCTION` | bool | `0` | debug | Trace function calls. |
| `FASTVIDEO_TRACE_ACTIVATIONS` | bool | `0` | debug | Enable activation trace hooks. |
| `FASTVIDEO_TRACE_LAYERS` | str | `""` | debug | Regex filter for traced module names. Empty means all. |
| `FASTVIDEO_TRACE_STATS` | str | `abs_mean,sum` | debug | Comma-separated activation statistics dumped for each output tensor. |
| `FASTVIDEO_TRACE_OUTPUT` | str | `/tmp/fv_trace_<pid>.jsonl` | debug | JSONL path for activation traces. The literal &lt;pid&gt; is replaced at runtime. |
| `FASTVIDEO_TRACE_STEPS` | str | `""` | debug | Comma-separated denoising step indices. Empty means all. |
| `FASTVIDEO_CFG_GATE_STEP` | float | `1.0` | sampling | CFG gating fraction in [0, 1]. Steps before len(timesteps) \* X run the conditional and unconditional forwards; later steps reuse the cached difference. 1.0 disables gating. |
| Variable | Type | Default | Category | Description |
| ---------------------------------------------------------- | ------------------------------ | ----------------------------------------------------- | ----------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `FASTVIDEO_CONFIG_ROOT` | path | computed | path | Root directory for FastVideo configuration files, at runtime and at installation. Defaults to ~/.config/fastvideo, or $XDG_CONFIG_HOME/fastvideo when XDG_CONFIG_HOME is set. |
| `FASTVIDEO_CACHE_ROOT` | path | computed | path | Root directory for FastVideo cache files. Defaults to ~/.cache/fastvideo, or $XDG_CACHE_HOME/fastvideo when XDG_CACHE_HOME is set. |
| `FASTVIDEO_REASON1_WEIGHTS_PATH` | str | unset | path | Local path or Hugging Face id of Reason1 weights to load instead of the checkpoint's own. |
| `FASTVIDEO_HOST_IP` | str | `""` | distributed | IP address of this node when the node has several network interfaces. Set it on each node for multi-node inference. |
| `FASTVIDEO_LOOPBACK_IP` | str | `""` | distributed | Loopback IP address to use instead of the detected one. |
| `FASTVIDEO_RAY_PER_WORKER_GPUS` | float | `1.0` | distributed | GPUs per Ray worker. A fraction lets Ray schedule several actors on one GPU, so other actors can share the GPUs with FastVideo. |
| `FASTVIDEO_NCCL_SO_PATH` | str | unset | distributed | Path to the NCCL library file. Needed because the nccl&gt;=2.19 that PyTorch ships has a bug (https://github.com/NVIDIA/nccl/issues/1234). |
| `FASTVIDEO_HCCL_SO_PATH` | str | unset | distributed | Path to the HCCL library file on Ascend NPUs. Deprecated names: `HCCL_SO_PATH`. |
| `FASTVIDEO_WORKER_MULTIPROC_METHOD` | one of spawn, fork, forkserver | `spawn` | distributed | Multiprocessing start method for worker processes. |
| `FASTVIDEO_ULYSSES_A2A` | one of off, auto | `off` | distributed | Sequence-parallel all-to-all backend. off uses the NCCL path in DistributedAutograd.AllToAll4D. auto uses the fused NVLink kernel when the group is a load-store accessible mesh of 2, 4, 6, or 8 ranks in eager execution, and the NCCL path otherwise. |
| `FASTVIDEO_CONFIGURE_LOGGING` | bool | `1` | logging | Configure logging at import. When true, FastVideo uses its default logging configuration or the file in FASTVIDEO_LOGGING_CONFIG_PATH. |
| `FASTVIDEO_LOGGING_CONFIG_PATH` | str | unset | logging | Path to a JSON logging configuration file. |
| `FASTVIDEO_LOGGING_LEVEL` | str | `INFO` | logging | Default logging level. |
| `FASTVIDEO_LOGGING_PREFIX` | str | `""` | logging | Prefix prepended to every log message. |
| `FASTVIDEO_STAGE_LOGGING` | bool | `0` | logging | Log the time that each pipeline stage takes. |
| `FASTVIDEO_ATTENTION_BACKEND` | str | unset | attention | Attention backend, as an AttentionBackendEnum name such as TORCH_SDPA, FLASH_ATTN, VIDEO_SPARSE_ATTN, SAGE_ATTN, or SAGE_ATTN_THREE. FastVideoArgs uses it when FastVideoArgs.attention_backend is unset. |
| `FASTVIDEO_FA4` | bool | `0` | attention | The FLASH_ATTN backend uses FlashAttention-4 (flash_attn.cute) instead of FA3 or FA2. |
| `FASTVIDEO_MINIMAX_H3_FA4_PACKED_VARLEN` | bool | `0` | attention | MiniMax-H3 dense DiT self-attention uses the FlashAttention-4 packed-varlen entry point. This changes the floating-point reduction order, so it is an inference-only opt-in. |
| `FASTVIDEO_VSA_SM100A` | bool | `0` | attention | VIDEO_SPARSE_ATTN_H3 sends no-grad tile-64 forwards to the data-center Blackwell (sm_100a) kernel. fastvideo-kernel reads the same variable with the same rule. |
| `FASTVIDEO_NVFP4_FA4` | bool | `0` | attention | FlashAttention-4 quantizes Q and K to NVFP4. An explicit nvfp4_fa4 attention implementation argument takes precedence. |
| `FASTVIDEO_DISABLE_ATTENTION_COMPILE` | bool | `1` | attention | Keep attention forward out of torch.compile graphs (torch.compiler.disable). Set it to 0 to let attention constructed under that setting be traced. Setting it explicitly to true also blocks regional compile. |
| `FASTVIDEO_MLX_WINDOW` | int | `0` | attention | MLX FastWan windowed attention size in tokens. 0 uses full attention. |
| `FASTVIDEO_MLX_WINDOW_SINK` | int | `0` | attention | Number of sink tokens that MLX windowed attention always attends to. |
| `FASTVIDEO_INFERENCE_TORCH_COMPILE` | bool | `0` | performance | Compile each DiT transformer block with fullgraph torch.compile at inference. Same as FastVideoArgs.inference_torch_compile=True. |
| `FASTVIDEO_VAE_PARALLEL_DECODE` | bool | `0` | performance | MiniMax-H3 VAE decode splits its temporal chunks across the sequence-parallel ranks instead of running serially on the output rank. Same as FastVideoArgs.vae_parallel_decode=True. |
| `FASTVIDEO_VAE_PARALLEL_ENCODE` | bool | `0` | performance | MiniMax-H3 reference-video VAE encode splits its temporal chunks across the sequence-parallel ranks. Same as FastVideoArgs.vae_parallel_encode=True. |
| `FASTVIDEO_VAE_PARALLEL_DECODE_STRATEGY` | str | unset | performance | Collective that moves chunks in parallel VAE decode: gather (used when unset) or all_gather. |
| `FASTVIDEO_MINIMAX_H3_FUSIONS` | str | `""` | performance | MiniMax-H3 inference-only Triton fusions: all, 1, or a comma-separated subset of modulate,qknorm_rope,swiglu. Empty, 0, or none keeps the eager implementation. |
| `FASTVIDEO_FSDP2_AUTOWRAP` | bool | `0` | performance | FSDP2 shards modules by parameter count instead of the model's shard conditions. Not supported by self-forcing distillation. |
| `FASTVIDEO_FSDP2_MIN_PARAMS` | int | `10000000` | performance | Minimum parameter count of a module that FASTVIDEO_FSDP2_AUTOWRAP shards. |
| `FASTVIDEO_MLX_COMPILE` | bool | `0` | performance | Compile the MLX DiT forward with mx.compile. |
| `FASTVIDEO_MLX_FAST_NORM` | bool | `0` | performance | Use MLX fast normalization kernels. |
| `FASTVIDEO_MLX_DQ_GEMM` | str | `1` | performance | MLX dequantized GEMM for affine-quantized weights: 0 turns it off, 1 uses the measured minimum row count, and an integer sets the minimum row count. |
| `FASTVIDEO_LTX2_VAE_CHANNELS_LAST_3D` | bool | `1` | performance | LTX-2 VAE uses the channels_last_3d memory format. |
| `FASTVIDEO_LTX2_DISABLE_AUDIO_AUTOCAST` | bool | `1` | performance | LTX-2 audio decoding runs without CUDA autocast. Deprecated names: `LTX2_DISABLE_AUDIO_AUTOCAST`. |
| `FASTVIDEO_FLUX2_DISABLE_BF16_REDUCED_PRECISION_REDUCTION` | bool | `0` | performance | Flux denoising disables reduced-precision reductions in bf16 matmuls, which tightens accumulation for the 4-step Klein model. |
| `FASTVIDEO_FFMPEG_BIN` | str | `ffmpeg` | output | ffmpeg executable used to save video with audio. |
| `FASTVIDEO_VIDEO_CODEC` | str | `libx264` | output | ffmpeg video codec for saved videos. |
| `FASTVIDEO_NVENC_PRESET` | str | `p1` | output | NVENC preset when the codec is an \*_nvenc codec. |
| `FASTVIDEO_NVENC_TUNE` | str | `ull` | output | NVENC tune option. |
| `FASTVIDEO_NVENC_RC` | str | `constqp` | output | NVENC rate-control mode. |
| `FASTVIDEO_NVENC_QP` | str | `28` | output | NVENC quantization parameter. |
| `FASTVIDEO_NVENC_BF` | str | `0` | output | NVENC number of B-frames. |
| `FASTVIDEO_X264_PRESET` | str | `ultrafast` | output | x264 preset for non-NVENC codecs. |
| `FASTVIDEO_OUTPUT_PIX_FMT` | str | `yuv420p` | output | ffmpeg pixel format for saved videos. |
| `FASTVIDEO_NVTX_PROFILE` | bool | `0` | profiling | Emit NVTX ranges for external profilers such as Nsight Systems. |
| `FASTVIDEO_TORCH_PROFILER_DIR` | path | unset | profiling | Enables the torch profiler and sets the directory for its traces. Must be an absolute path. |
| `FASTVIDEO_TORCH_PROFILER_RECORD_SHAPES` | bool | `0` | profiling | Torch profiler records shapes. |
| `FASTVIDEO_TORCH_PROFILER_WITH_PROFILE_MEMORY` | bool | `0` | profiling | Torch profiler profiles memory. |
| `FASTVIDEO_TORCH_PROFILER_WITH_STACK` | bool | `0` | profiling | Torch profiler captures stacks. Costs about 1.5x runtime and 1.4x trace size. |
| `FASTVIDEO_TORCH_PROFILER_WITH_FLOPS` | bool | `0` | profiling | Torch profiler profiles FLOPs. |
| `FASTVIDEO_TORCH_PROFILE_REGIONS` | str | `""` | profiling | Comma-separated profiler regions to record. The torch profiler requires at least one region. |
| `FASTVIDEO_TRACE_ACTIVATIONS` | bool | `0` | debug | Enable activation trace hooks. |
| `FASTVIDEO_TRACE_LAYERS` | str | `""` | debug | Regex filter for traced module names. Empty means all. |
| `FASTVIDEO_TRACE_STATS` | str | `abs_mean,sum` | debug | Comma-separated activation statistics dumped for each output tensor. |
| `FASTVIDEO_TRACE_OUTPUT` | str | `/tmp/fv_trace_<pid>.jsonl` | debug | JSONL path for activation traces. The literal &lt;pid&gt; is replaced at runtime. |
| `FASTVIDEO_TRACE_STEPS` | str | `""` | debug | Comma-separated denoising step indices. Empty means all. |
| `FASTVIDEO_H3_VSA_PROBE` | str | unset | debug | Output directory for the VSA-H3 attention-mass probe, which writes one .pt file per step, layer, and rank. Keeps the model out of regional compile. |
| `FASTVIDEO_LTX2_GEMMA_LOG` | str | `""` | debug | Log file for LTX-2 Gemma text-encoder hidden states, used by parity tests. Deprecated names: `LTX2_FASTVIDEO_GEMMA_LOG`. |
| `FASTVIDEO_COSMOS25_LOG_KNOBS` | bool | `0` | debug | Log the Cosmos 2.5 latent-preparation conditioning inputs. |
| `FASTVIDEO_CFG_GATE_STEP` | float | `1.0` | sampling | CFG gating fraction in [0, 1]. Steps before len(timesteps) \* X run the conditional and unconditional forwards; later steps reuse the cached difference. 1.0 disables gating. |
| `FASTVIDEO_LTX2_USE_DISTILLED_SIGMAS` | bool | `1` | sampling | LTX-2 uses the distilled sigma schedule when FastVideoArgs.ltx2_use_distilled_sigmas is also true. Deprecated names: `LTX2_USE_DISTILLED_SIGMAS`. |
| `FASTVIDEO_EVAL_CACHE` | path | computed | eval | Cache directory for evaluation models and datasets. Defaults to $FASTVIDEO_CACHE_ROOT/eval. |
| `FASTVIDEO_PHYSICS_IQ_BUCKET_URL` | str | `https://storage.googleapis.com/physics-iq-benchmark` | eval | Base URL of the Physics-IQ benchmark bucket. |
| `FASTVIDEO_VBENCH_FULL_INFO_JSON` | str | unset | eval | Path to VBench_full_info.json, used instead of the vendored copy. Deprecated names: `VBENCH_FULL_INFO_JSON`. |
| `FASTVIDEO_FVD_REF_FEATURES` | str | unset | eval | Cached reference-feature file for the FVD metric. |
| `FASTVIDEO_FAD_REF_FEATURES` | str | unset | eval | Cached reference-feature file for the audio Frechet distance metric. |
| `FASTVIDEO_TEST_LTX2_OVERFIT_DATA_DIR` | str | `data/cats` | test | Raw data directory for preprocess_ltx2_overfit.py. Deprecated names: `LTX2_OVERFIT_DATA_DIR`. |
| `FASTVIDEO_TEST_LTX2_OVERFIT_CAPTION_JSON` | str | `videos2caption_1_sample.json` | test | Caption file, relative to the raw data directory. Deprecated names: `LTX2_OVERFIT_CAPTION_JSON`. |
| `FASTVIDEO_TEST_LTX2_OVERFIT_VIDEO_SUBDIR` | str | `video` | test | Video subdirectory, relative to the raw data directory. Deprecated names: `LTX2_OVERFIT_VIDEO_SUBDIR`. |
| `FASTVIDEO_TEST_LTX2_OVERFIT_OUTPUT_DIR` | str | `data/ltx2_overfit_preprocessed` | test | Output directory for preprocess_ltx2_overfit.py. Deprecated names: `LTX2_OVERFIT_OUTPUT_DIR`. |
| `FASTVIDEO_TEST_LTX2_OVERFIT_MODEL` | str | `FastVideo/LTX2-Distilled-Diffusers` | test | Model repository whose encoders preprocess_ltx2_overfit.py uses. Deprecated names: `LTX2_OVERFIT_MODEL`. |
| `FASTVIDEO_TEST_LTX2_OVERFIT_NUM_COPIES` | int | `4` | test | Number of copies of the overfit sample in the parquet file. Deprecated names: `LTX2_OVERFIT_NUM_COPIES`. |
| `FASTVIDEO_TEST_KANDINSKY5_OVERFIT_DATA_DIR` | str | `data/kandinsky5_overfit` | test | Raw data directory for preprocess_kandinsky5_overfit.py. Deprecated names: `KANDINSKY5_OVERFIT_DATA_DIR`. |
| `FASTVIDEO_TEST_KANDINSKY5_OVERFIT_OUTPUT_DIR` | str | `data/kandinsky5_overfit_preprocessed` | test | Output directory for preprocess_kandinsky5_overfit.py. Deprecated names: `KANDINSKY5_OVERFIT_OUTPUT_DIR`. |
Variables that FastVideo no longer reads; setting one logs a warning:
| Deprecated variable | Reason |
| ----------------------------------------- | ---------------- |
| `FASTVIDEO_TARGET_DEVICE` | no code reads it |
| `FASTVIDEO_USE_PRECOMPILED` | no code reads it |
| `FASTVIDEO_RINGBUFFER_WARNING_INTERVAL` | no code reads it |
| `FASTVIDEO_ENGINE_ITERATION_TIMEOUT_S` | no code reads it |
| `FASTVIDEO_SERVER_DEV_MODE` | no code reads it |
| `FASTVIDEO_TEST_DYNAMO_FULLGRAPH_CAPTURE` | no code reads it |
| `FASTVIDEO_TRACE_FUNCTION` | no code reads it |
<!-- END GENERATED ENV TABLE -->
examples/train/configs/fine_tuning/kandinsky5/README.md
+1 -1
View File
@@ -27,7 +27,7 @@ layout) with:
```bash
# from the repo root; reads data/kandinsky5_overfit and writes
# data/kandinsky5_overfit_preprocessed by default -- override with
# KANDINSKY5_OVERFIT_DATA_DIR / KANDINSKY5_OVERFIT_OUTPUT_DIR env vars
# FASTVIDEO_TEST_KANDINSKY5_OVERFIT_DATA_DIR / FASTVIDEO_TEST_KANDINSKY5_OVERFIT_OUTPUT_DIR env vars
python -m fastvideo.pipelines.preprocess.preprocess_kandinsky5_overfit
```
examples/train/configs/fine_tuning/ltx2_3/nvfp4_qat_t2v.yaml
+1 -1
View File
@@ -32,7 +32,7 @@
# (see docs/training/data_preprocess.md;
# fastvideo/pipelines/preprocess/preprocess_ltx2_overfit.py is a minimal
# LTX-2 reference to adapt — set
# LTX2_OVERFIT_MODEL=FastVideo/LTX-2.3-Distilled-Diffusers). Data paths,
# FASTVIDEO_TEST_LTX2_OVERFIT_MODEL=FastVideo/LTX-2.3-Distilled-Diffusers). Data paths,
# step counts, and learning rate below are placeholders you must adapt
# to your workload.
#
examples/train/configs/fine_tuning/ltx2_3/t2v.yaml
+1 -1
View File
@@ -12,7 +12,7 @@
# (see docs/training/data_preprocess.md;
# fastvideo/pipelines/preprocess/preprocess_ltx2_overfit.py is a minimal
# LTX-2 reference to adapt — set
# LTX2_OVERFIT_MODEL=FastVideo/LTX-2.3-Distilled-Diffusers).
# FASTVIDEO_TEST_LTX2_OVERFIT_MODEL=FastVideo/LTX-2.3-Distilled-Diffusers).
#
# Run:
# NUM_GPUS=4 bash examples/train/run.sh \
examples/train/configs/overfit_ltx2_3_t2v.yaml
+2 -2
View File
@@ -8,8 +8,8 @@
#
# Preprocess data first (writes data/ltx2_3_overfit_preprocessed):
# CUDA_VISIBLE_DEVICES=0 \
# LTX2_OVERFIT_MODEL=FastVideo/LTX-2.3-Distilled-Diffusers \
# LTX2_OVERFIT_OUTPUT_DIR=data/ltx2_3_overfit_preprocessed \
# FASTVIDEO_TEST_LTX2_OVERFIT_MODEL=FastVideo/LTX-2.3-Distilled-Diffusers \
# FASTVIDEO_TEST_LTX2_OVERFIT_OUTPUT_DIR=data/ltx2_3_overfit_preprocessed \
# python fastvideo/pipelines/preprocess/preprocess_ltx2_overfit.py
#
# Run:
fastvideo/attention/backends/flash_attn.py
+2 -1
View File
@@ -5,6 +5,7 @@ import torch
import torch.nn.functional as F
from dataclasses import dataclass
import fastvideo.envs as envs
from fastvideo.attention.utils.flash_attn_default import (
fa_version,
flash_attn_func_compilable,
@@ -264,7 +265,7 @@ class FlashAttentionImpl(AttentionImpl):
# keeps its previous semantics.
nvfp4_fa4 = extra_impl_args.get("nvfp4_fa4")
if nvfp4_fa4 is None:
nvfp4_fa4 = os.environ.get("FASTVIDEO_NVFP4_FA4", "0") == "1"
nvfp4_fa4 = envs.FASTVIDEO_NVFP4_FA4.get()
self.nvfp4_fa4 = bool(nvfp4_fa4)
if self.nvfp4_fa4:
cap = torch.cuda.get_device_capability()
fastvideo/attention/backends/video_sparse_attn_h3_probe.py
+3 -1
View File
@@ -23,12 +23,14 @@ import os
import torch
import fastvideo.envs as envs
_FRACS = (0.05, 0.10, 0.125, 0.25, 0.50, 0.75)
_TRUE_ROWS = 128 # sampled query rows per layer for the token-true check
def probe_enabled() -> str | None:
return os.environ.get("FASTVIDEO_H3_VSA_PROBE") or None
return envs.FASTVIDEO_H3_VSA_PROBE.get() or None
@torch.no_grad()
fastvideo/attention/layer.py
+3 -6
View File
@@ -1,11 +1,11 @@
# SPDX-License-Identifier: Apache-2.0
import os
from functools import wraps
import torch
import torch.nn as nn
import fastvideo.envs as envs
from fastvideo.attention.selector import backend_name_to_enum, get_attn_backend
from fastvideo.distributed.communication_op import (sequence_model_parallel_all_gather,
sequence_model_parallel_all_to_all_4D)
@@ -23,10 +23,7 @@ def _attention_compile_disabled() -> bool:
``torch.compiler.disable``). Set ``FASTVIDEO_DISABLE_ATTENTION_COMPILE=0``
to let attention instances constructed under that environment be traced.
"""
val = os.environ.get("FASTVIDEO_DISABLE_ATTENTION_COMPILE")
if val is None:
return True
return val.strip().lower() not in ("0", "false", "no", "off", "")
return envs.FASTVIDEO_DISABLE_ATTENTION_COMPILE.get()
def _attention_compile_explicitly_disabled() -> bool:
@@ -35,7 +32,7 @@ def _attention_compile_explicitly_disabled() -> bool:
Regional compile can override the historical default for one loaded
transformer, but it must still honor an explicit debugging escape hatch.
"""
return "FASTVIDEO_DISABLE_ATTENTION_COMPILE" in os.environ and _attention_compile_disabled()
return envs.FASTVIDEO_DISABLE_ATTENTION_COMPILE.is_set() and _attention_compile_disabled()
def _maybe_compiler_disable(fn):
fastvideo/attention/selector.py
+2 -3
View File
@@ -1,7 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# Adapted from vllm: https://github.com/vllm-project/vllm/blob/v0.7.3/vllm/attention/selector.py
import os
from collections.abc import Generator
from contextlib import contextmanager
from contextvars import ContextVar
@@ -15,7 +14,7 @@ import fastvideo.envs as envs
from fastvideo.attention.backends.abstract import AttentionBackend
from fastvideo.logger import init_logger
from fastvideo.platforms import AttentionBackendEnum
from fastvideo.utils import STR_BACKEND_ENV_VAR, resolve_obj_by_qualname
from fastvideo.utils import resolve_obj_by_qualname
logger = init_logger(__name__)
@@ -65,7 +64,7 @@ def get_env_variable_attn_backend() -> AttentionBackendEnum | None:
* _Backend enum value if an override is specified
* None otherwise
'''
backend_name = os.environ.get(STR_BACKEND_ENV_VAR)
backend_name = envs.FASTVIDEO_ATTENTION_BACKEND.get()
return (None if backend_name is None else backend_name_to_enum(backend_name))
fastvideo/benchmarks/mlx_fastwan_bench.py
+3 -2
View File
@@ -46,6 +46,7 @@ from examples.inference.basic.mlx_wan_prompt_to_video import (
encode_prompt,
make_rotary_embeddings,
)
import fastvideo.envs as envs
from fastvideo.mlx_runtime.memory import add_memory_limit_args, apply_memory_limits, cleanup_mlx
# The highest-fidelity cell; used as the default SSIM reference when no external
@@ -524,8 +525,8 @@ def _generate_cell(
"peak_gib": max(load_peak, denoise_peak) / (1024**3),
"quantization": quantization or "none",
"compute_dtype": base_dtype,
"compile": os.environ.get("FASTVIDEO_MLX_COMPILE", "0") == "1",
"fast_norm": os.environ.get("FASTVIDEO_MLX_FAST_NORM", "0") == "1",
"compile": envs.FASTVIDEO_MLX_COMPILE.get(),
"fast_norm": envs.FASTVIDEO_MLX_FAST_NORM.get(),
"mlx_memory_limit_gib": args.mlx_memory_limit_gib,
"mlx_cache_limit_gib": args.mlx_cache_limit_gib,
"mlx_disable_cache": args.mlx_disable_cache,
fastvideo/distributed/device_communicators/pyhccl_wrapper.py
+1 -1
View File
@@ -144,7 +144,7 @@ class HCCLLibrary:
"Otherwise, the hccl library might not exist, be corrupted "
"or it does not support the current platform %s. "
"If you already have the library, please set the "
"environment variable HCCL_SO_PATH"
"environment variable FASTVIDEO_HCCL_SO_PATH"
" to point to the correct hccl library path.", so_file, platform.platform())
raise e
fastvideo/distributed/parallel_state.py
+3 -4
View File
@@ -40,7 +40,6 @@ import torch.distributed
import torch.distributed as dist
from torch.distributed import Backend, ProcessGroup, ReduceOp
import fastvideo.envs as envs
from fastvideo.distributed.device_communicators.base_device_communicator import (DeviceCommunicatorBase)
from fastvideo.distributed.device_communicators.cpu_communicator import (CpuCommunicator)
from fastvideo.distributed.utils import StatelessProcessGroup
@@ -759,7 +758,7 @@ def init_distributed_environment(
if local_rank == -1:
# local rank not set, this usually happens in single-node
# setting, where we can use rank as local rank
local_rank = envs.LOCAL_RANK.get() if distributed_init_method == "env://" else rank
local_rank = int(os.environ.get("LOCAL_RANK", "0")) if distributed_init_method == "env://" else rank
global _WORLD
if _WORLD is None:
ranks = list(range(torch.distributed.get_world_size()))
@@ -884,9 +883,9 @@ def get_local_torch_device() -> torch.device:
"""Return the torch device for the current rank."""
from fastvideo.platforms import current_platform
if current_platform.is_npu():
device = torch.device(f"npu:{envs.LOCAL_RANK.get()}")
device = torch.device(f"npu:{int(os.environ.get('LOCAL_RANK', '0'))}")
elif current_platform.is_cuda_alike() or current_platform.is_cuda():
device = torch.device(f"cuda:{envs.LOCAL_RANK.get()}")
device = torch.device(f"cuda:{int(os.environ.get('LOCAL_RANK', '0'))}")
else:
device = torch.device("mps")
return device
fastvideo/entrypoints/streaming/worker.py
+2 -2
View File
@@ -50,9 +50,9 @@ def worker_main(
``VideoGenerator`` construction + generation happens here, not in
the parent's event loop.
"""
import os
import fastvideo.envs as envs
os.environ["CUDA_VISIBLE_DEVICES"] = str(gpu_id)
envs.set_external("CUDA_VISIBLE_DEVICES", str(gpu_id))
try:
from fastvideo import VideoGenerator
fastvideo/entrypoints/video_generator.py
+11 -10
View File
@@ -26,6 +26,7 @@ import torch
import torchvision
from einops import rearrange
import fastvideo.envs as envs
from fastvideo.api.compat import (
REQUEST_BATCH_EXTRA_PASSTHROUGH_FIELDS,
expand_request_prompt_batch,
@@ -200,7 +201,7 @@ class VideoGenerator:
if kwargs.pop("nvfp4_fa4", False):
import os
os.environ["FASTVIDEO_NVFP4_FA4"] = "1"
os.environ.setdefault("CUTE_DSL_ENABLE_TVM_FFI", "1")
envs.setdefault_external("CUTE_DSL_ENABLE_TVM_FFI", "1")
typed_config = kwargs.pop("config", None)
if typed_config is not None:
if model_path is not None:
@@ -1191,7 +1192,7 @@ class VideoGenerator:
sample_rate: int,
) -> bool:
"""Encode video+audio using ffmpeg via rawvideo stdin + WAV input."""
ffmpeg_bin = shutil.which(os.getenv("FASTVIDEO_FFMPEG_BIN", "ffmpeg"))
ffmpeg_bin = shutil.which(envs.FASTVIDEO_FFMPEG_BIN.get())
if ffmpeg_bin is None:
logger.warning("ffmpeg not found; cannot use ffmpeg pipe save.")
return False
@@ -1201,7 +1202,7 @@ class VideoGenerator:
height = int(frames[0].shape[0])
width = int(frames[0].shape[1])
codec = os.getenv("FASTVIDEO_VIDEO_CODEC", "libx264")
codec = envs.FASTVIDEO_VIDEO_CODEC.get()
try:
with tempfile.TemporaryDirectory() as tmpdir:
@@ -1236,23 +1237,23 @@ class VideoGenerator:
if codec.endswith("_nvenc"):
nvenc_options = [
("preset", os.getenv("FASTVIDEO_NVENC_PRESET", "p1")),
("tune", os.getenv("FASTVIDEO_NVENC_TUNE", "ull")),
("rc", os.getenv("FASTVIDEO_NVENC_RC", "constqp")),
("qp", os.getenv("FASTVIDEO_NVENC_QP", "28")),
("bf", os.getenv("FASTVIDEO_NVENC_BF", "0")),
("preset", envs.FASTVIDEO_NVENC_PRESET.get()),
("tune", envs.FASTVIDEO_NVENC_TUNE.get()),
("rc", envs.FASTVIDEO_NVENC_RC.get()),
("qp", envs.FASTVIDEO_NVENC_QP.get()),
("bf", envs.FASTVIDEO_NVENC_BF.get()),
]
for option_name, option_value in nvenc_options:
if cls._ffmpeg_encoder_supports_option(ffmpeg_bin, codec, option_name):
cmd += [f"-{option_name}", option_value]
else:
cmd += ["-preset", os.getenv("FASTVIDEO_X264_PRESET", "ultrafast")]
cmd += ["-preset", envs.FASTVIDEO_X264_PRESET.get()]
cmd += [
"-c:a",
"aac",
"-pix_fmt",
os.getenv("FASTVIDEO_OUTPUT_PIX_FMT", "yuv420p"),
envs.FASTVIDEO_OUTPUT_PIX_FMT.get(),
# Audio and video decoders may produce different durations.
# Preserve every video frame and pad/trim audio to match.
"-af",
fastvideo/envs.py
+219 -55
View File
@@ -12,6 +12,7 @@ The policy for environment variables is in ``docs/contributing/env_vars.md``,
and ``fastvideo/tests/contract/test_env_policy.py`` enforces it.
"""
import logging
import os
from collections.abc import Callable, Iterator
from contextlib import contextmanager
@@ -24,12 +25,23 @@ S = TypeVar("S", str, str | None)
POLICY_DOC = "docs/contributing/env_vars.md"
# Allowed values of EnvField.category.
CATEGORIES = ("build", "path", "distributed", "external", "logging", "attention", "performance", "profiling", "debug",
"sampling")
CATEGORIES = ("path", "distributed", "logging", "attention", "performance", "output", "profiling", "debug", "sampling",
"eval", "test")
_TRUE_VALUES = frozenset({"1", "true", "yes", "on"})
_FALSE_VALUES = frozenset({"0", "false", "no", "off", ""})
# fastvideo.logger imports this module, so warnings go through the standard
# logging module; the "fastvideo" logger configuration still applies.
_logger = logging.getLogger(__name__)
_warned_messages: set[str] = set()
def _warn_once(message: str) -> None:
if message not in _warned_messages:
_warned_messages.add(message)
_logger.warning(message)
class EnvVarError(ValueError):
"""An environment variable holds a value that its registered type rejects."""
@@ -39,16 +51,20 @@ class EnvField(Generic[T]):
"""One registered environment variable: its type, default, category, and description.
``default`` is either the value itself or a zero-argument function that
computes it on each read while the variable is unset.
computes it on each read while the variable is unset. ``deprecated_names``
lists earlier names of a renamed variable; they are read, with a warning,
only when the variable itself is unset.
"""
type_name = ""
def __init__(self, default: T | Callable[[], T], *, category: str, doc: str) -> None:
def __init__(self, default: T | Callable[[], T], *, category: str, doc: str,
deprecated_names: tuple[str, ...] = ()) -> None:
self.name = "" # Set by _register_fields() from the module attribute name.
self.default = default
self.category = category
self.doc = doc
self.deprecated_names = deprecated_names
def parse(self, raw: str) -> T:
raise NotImplementedError
@@ -56,40 +72,55 @@ class EnvField(Generic[T]):
def format(self, value: T) -> str:
return str(value)
def _read_raw(self) -> tuple[str, str] | None:
"""Return (name, raw value) from the variable or its first set deprecated name."""
raw = os.environ.get(self.name)
if raw is not None:
return self.name, raw
for old_name in self.deprecated_names:
raw = os.environ.get(old_name)
if raw is not None:
_warn_once(f"{old_name} is deprecated and will be removed in the next minor release; "
f"set {self.name} instead.")
return old_name, raw
return None
def get(self) -> T:
"""Return the parsed value, or the default when the variable is unset."""
raw = os.environ.get(self.name)
if raw is None:
found = self._read_raw()
if found is None:
return self.default() if callable(self.default) else self.default
name, raw = found
try:
return self.parse(raw)
except ValueError as exc:
raise EnvVarError(f"Invalid value {raw!r} for {self.name}: {exc}. See {POLICY_DOC}.") from None
raise EnvVarError(f"Invalid value {raw!r} for {name}: {exc}. See {POLICY_DOC}.") from None
def is_set(self) -> bool:
return self.name in os.environ
return any(name in os.environ for name in (self.name, *self.deprecated_names))
def set(self, value: T) -> None:
os.environ[self.name] = self.format(value)
def clear(self) -> None:
os.environ.pop(self.name, None)
for name in (self.name, *self.deprecated_names):
os.environ.pop(name, None)
@contextmanager
def override(self, value: T | None) -> Iterator[None]:
"""Set the variable, or unset it when ``value`` is None, and restore the previous value on exit."""
previous = os.environ.get(self.name)
if value is None:
self.clear()
else:
"""Set the variable, or unset it when ``value`` is None, and restore the previous values on exit."""
previous = {name: os.environ.get(name) for name in (self.name, *self.deprecated_names)}
self.clear()
if value is not None:
self.set(value)
try:
yield
finally:
if previous is None:
self.clear()
else:
os.environ[self.name] = previous
for name, old_value in previous.items():
if old_value is None:
os.environ.pop(name, None)
else:
os.environ[name] = old_value
def __bool__(self) -> bool:
raise TypeError(f"Use envs.{self.name}.get() to read {self.name}.")
@@ -171,20 +202,25 @@ def get_default_config_root() -> str:
)
# ================== Installation ==================
# Variables that other tools read (CUDA, NCCL, PyTorch, launchers) are written
# only through these helpers. The contract test checks each name against its
# write allowlist.
def set_external(name: str, value: str) -> None:
"""Set an environment variable that another tool reads."""
os.environ[name] = value
def setdefault_external(name: str, value: str) -> None:
"""Set an environment variable that another tool reads, unless it is already set."""
os.environ.setdefault(name, value)
def unset_external(name: str) -> None:
"""Remove an environment variable that another tool reads."""
os.environ.pop(name, None)
FASTVIDEO_TARGET_DEVICE = EnvStr("cuda",
category="build",
doc="Target device of FastVideo: cuda, rocm, neuron, cpu, or openvino.")
MAX_JOBS = EnvStr(None,
category="build",
doc="Maximum number of parallel compilation jobs. Defaults to the number of CPUs.")
NVCC_THREADS = EnvStr(None,
category="build",
doc="Number of nvcc threads. When set, MAX_JOBS is reduced to avoid oversubscribing the CPU.")
FASTVIDEO_USE_PRECOMPILED = EnvBool(False, category="build", doc="Use precompiled binaries (*.so).")
CMAKE_BUILD_TYPE = EnvStr(None, category="build", doc="CMake build type: Debug, Release, or RelWithDebInfo.")
VERBOSE = EnvBool(False, category="build", doc="Print verbose logs during installation.")
# ================== Paths ==================
@@ -198,6 +234,10 @@ FASTVIDEO_CACHE_ROOT = EnvPath(
category="path",
doc="Root directory for FastVideo cache files. "
"Defaults to ~/.cache/fastvideo, or $XDG_CACHE_HOME/fastvideo when XDG_CACHE_HOME is set.")
FASTVIDEO_REASON1_WEIGHTS_PATH = EnvStr(None,
category="path",
doc="Local path or Hugging Face id of Reason1 weights to load instead of the "
"checkpoint's own.")
# ================== Distributed ==================
@@ -214,18 +254,15 @@ FASTVIDEO_RAY_PER_WORKER_GPUS = EnvFloat(
category="distributed",
doc="GPUs per Ray worker. A fraction lets Ray schedule several actors on one GPU, so other actors can share "
"the GPUs with FastVideo.")
FASTVIDEO_RINGBUFFER_WARNING_INTERVAL = EnvInt(60,
category="distributed",
doc="Seconds between warnings while the ring buffer is full.")
FASTVIDEO_NCCL_SO_PATH = EnvStr(
None,
category="distributed",
doc="Path to the NCCL library file. Needed because the nccl>=2.19 that PyTorch ships has a bug "
"(https://github.com/NVIDIA/nccl/issues/1234).")
HCCL_SO_PATH = EnvStr(None, category="distributed", doc="Path to the HCCL library file on Ascend NPUs.")
FASTVIDEO_ENGINE_ITERATION_TIMEOUT_S = EnvInt(60,
category="distributed",
doc="Timeout in seconds for each engine iteration.")
FASTVIDEO_HCCL_SO_PATH = EnvStr(None,
category="distributed",
doc="Path to the HCCL library file on Ascend NPUs.",
deprecated_names=("HCCL_SO_PATH", ))
FASTVIDEO_WORKER_MULTIPROC_METHOD = EnvChoice("spawn",
choices=("spawn", "fork", "forkserver"),
category="distributed",
@@ -238,18 +275,6 @@ FASTVIDEO_ULYSSES_A2A = EnvChoice(
"the fused NVLink kernel when the group is a load-store accessible mesh of 2, 4, 6, or 8 ranks in eager "
"execution, and the NCCL path otherwise.")
# ================== External variables ==================
# Variables that other tools set. They stay registered so that Ray copies them
# to its workers until the external-variable allowlist replaces them.
LD_LIBRARY_PATH = EnvStr(None,
category="external",
doc="Searched for the NCCL library when FASTVIDEO_NCCL_SO_PATH is unset.")
LOCAL_RANK = EnvInt(0,
category="external",
doc="Local rank of the process in a distributed run; selects the GPU device id.")
CUDA_VISIBLE_DEVICES = EnvStr(None, category="external", doc="Visible devices in a distributed run.")
# ================== Logging ==================
FASTVIDEO_CONFIGURE_LOGGING = EnvBool(
@@ -285,6 +310,22 @@ FASTVIDEO_VSA_SM100A = EnvBool(
category="attention",
doc="VIDEO_SPARSE_ATTN_H3 sends no-grad tile-64 forwards to the data-center Blackwell (sm_100a) kernel. "
"fastvideo-kernel reads the same variable with the same rule.")
FASTVIDEO_NVFP4_FA4 = EnvBool(
False,
category="attention",
doc="FlashAttention-4 quantizes Q and K to NVFP4. An explicit nvfp4_fa4 attention implementation argument "
"takes precedence.")
FASTVIDEO_DISABLE_ATTENTION_COMPILE = EnvBool(
True,
category="attention",
doc="Keep attention forward out of torch.compile graphs (torch.compiler.disable). Set it to 0 to let attention "
"constructed under that setting be traced. Setting it explicitly to true also blocks regional compile.")
FASTVIDEO_MLX_WINDOW = EnvInt(0,
category="attention",
doc="MLX FastWan windowed attention size in tokens. 0 uses full attention.")
FASTVIDEO_MLX_WINDOW_SINK = EnvInt(0,
category="attention",
doc="Number of sink tokens that MLX windowed attention always attends to.")
# ================== Performance ==================
@@ -316,7 +357,44 @@ FASTVIDEO_MINIMAX_H3_FUSIONS = EnvStr(
category="performance",
doc="MiniMax-H3 inference-only Triton fusions: all, 1, or a comma-separated subset of "
"modulate,qknorm_rope,swiglu. Empty, 0, or none keeps the eager implementation.")
FASTVIDEO_TEST_DYNAMO_FULLGRAPH_CAPTURE = EnvBool(True, category="debug", doc="Enable Dynamo fullgraph capture.")
FASTVIDEO_FSDP2_AUTOWRAP = EnvBool(False,
category="performance",
doc="FSDP2 shards modules by parameter count instead of the model's shard "
"conditions. Not supported by self-forcing distillation.")
FASTVIDEO_FSDP2_MIN_PARAMS = EnvInt(10000000,
category="performance",
doc="Minimum parameter count of a module that FASTVIDEO_FSDP2_AUTOWRAP shards.")
FASTVIDEO_MLX_COMPILE = EnvBool(False, category="performance", doc="Compile the MLX DiT forward with mx.compile.")
FASTVIDEO_MLX_FAST_NORM = EnvBool(False, category="performance", doc="Use MLX fast normalization kernels.")
FASTVIDEO_MLX_DQ_GEMM = EnvStr(
"1",
category="performance",
doc="MLX dequantized GEMM for affine-quantized weights: 0 turns it off, 1 uses the measured minimum row count, "
"and an integer sets the minimum row count.")
FASTVIDEO_LTX2_VAE_CHANNELS_LAST_3D = EnvBool(True,
category="performance",
doc="LTX-2 VAE uses the channels_last_3d memory format.")
FASTVIDEO_LTX2_DISABLE_AUDIO_AUTOCAST = EnvBool(True,
category="performance",
doc="LTX-2 audio decoding runs without CUDA autocast.",
deprecated_names=("LTX2_DISABLE_AUDIO_AUTOCAST", ))
FASTVIDEO_FLUX2_DISABLE_BF16_REDUCED_PRECISION_REDUCTION = EnvBool(
False,
category="performance",
doc="Flux denoising disables reduced-precision reductions in bf16 matmuls, which tightens accumulation for "
"the 4-step Klein model.")
# ================== Output encoding ==================
FASTVIDEO_FFMPEG_BIN = EnvStr("ffmpeg", category="output", doc="ffmpeg executable used to save video with audio.")
FASTVIDEO_VIDEO_CODEC = EnvStr("libx264", category="output", doc="ffmpeg video codec for saved videos.")
FASTVIDEO_NVENC_PRESET = EnvStr("p1", category="output", doc="NVENC preset when the codec is an *_nvenc codec.")
FASTVIDEO_NVENC_TUNE = EnvStr("ull", category="output", doc="NVENC tune option.")
FASTVIDEO_NVENC_RC = EnvStr("constqp", category="output", doc="NVENC rate-control mode.")
FASTVIDEO_NVENC_QP = EnvStr("28", category="output", doc="NVENC quantization parameter.")
FASTVIDEO_NVENC_BF = EnvStr("0", category="output", doc="NVENC number of B-frames.")
FASTVIDEO_X264_PRESET = EnvStr("ultrafast", category="output", doc="x264 preset for non-NVENC codecs.")
FASTVIDEO_OUTPUT_PIX_FMT = EnvStr("yuv420p", category="output", doc="ffmpeg pixel format for saved videos.")
# ================== Profiling ==================
@@ -341,10 +419,6 @@ FASTVIDEO_TORCH_PROFILE_REGIONS = EnvStr(
# ================== Debug ==================
FASTVIDEO_SERVER_DEV_MODE = EnvBool(False,
category="debug",
doc="Run the server in development mode with extra debugging endpoints.")
FASTVIDEO_TRACE_FUNCTION = EnvBool(False, category="debug", doc="Trace function calls.")
FASTVIDEO_TRACE_ACTIVATIONS = EnvBool(False, category="debug", doc="Enable activation trace hooks.")
FASTVIDEO_TRACE_LAYERS = EnvStr("", category="debug", doc="Regex filter for traced module names. Empty means all.")
FASTVIDEO_TRACE_STATS = EnvStr("abs_mean,sum",
@@ -354,6 +428,18 @@ FASTVIDEO_TRACE_OUTPUT = EnvStr("/tmp/fv_trace_<pid>.jsonl",
category="debug",
doc="JSONL path for activation traces. The literal <pid> is replaced at runtime.")
FASTVIDEO_TRACE_STEPS = EnvStr("", category="debug", doc="Comma-separated denoising step indices. Empty means all.")
FASTVIDEO_H3_VSA_PROBE = EnvStr(
None,
category="debug",
doc="Output directory for the VSA-H3 attention-mass probe, which writes one .pt file per step, layer, and rank. "
"Keeps the model out of regional compile.")
FASTVIDEO_LTX2_GEMMA_LOG = EnvStr("",
category="debug",
doc="Log file for LTX-2 Gemma text-encoder hidden states, used by parity tests.",
deprecated_names=("LTX2_FASTVIDEO_GEMMA_LOG", ))
FASTVIDEO_COSMOS25_LOG_KNOBS = EnvBool(False,
category="debug",
doc="Log the Cosmos 2.5 latent-preparation conditioning inputs.")
# ================== Sampling ==================
@@ -385,6 +471,84 @@ FASTVIDEO_CFG_GATE_STEP = EnvFloat(
category="sampling",
doc="CFG gating fraction in [0, 1]. Steps before len(timesteps) * X run the conditional and unconditional "
"forwards; later steps reuse the cached difference. 1.0 disables gating.")
FASTVIDEO_LTX2_USE_DISTILLED_SIGMAS = EnvBool(
True,
category="sampling",
doc="LTX-2 uses the distilled sigma schedule when FastVideoArgs.ltx2_use_distilled_sigmas is also true.",
deprecated_names=("LTX2_USE_DISTILLED_SIGMAS", ))
# ================== Evaluation ==================
FASTVIDEO_EVAL_CACHE = EnvPath(lambda: os.path.join(FASTVIDEO_CACHE_ROOT.get(), "eval"),
category="eval",
doc="Cache directory for evaluation models and datasets. Defaults to "
"$FASTVIDEO_CACHE_ROOT/eval.")
FASTVIDEO_PHYSICS_IQ_BUCKET_URL = EnvStr("https://storage.googleapis.com/physics-iq-benchmark",
category="eval",
doc="Base URL of the Physics-IQ benchmark bucket.")
FASTVIDEO_VBENCH_FULL_INFO_JSON = EnvStr(None,
category="eval",
doc="Path to VBench_full_info.json, used instead of the vendored copy.",
deprecated_names=("VBENCH_FULL_INFO_JSON", ))
FASTVIDEO_FVD_REF_FEATURES = EnvStr(None, category="eval", doc="Cached reference-feature file for the FVD metric.")
FASTVIDEO_FAD_REF_FEATURES = EnvStr(None,
category="eval",
doc="Cached reference-feature file for the audio Frechet distance metric.")
# ================== Tests ==================
FASTVIDEO_TEST_LTX2_OVERFIT_DATA_DIR = EnvStr("data/cats",
category="test",
doc="Raw data directory for preprocess_ltx2_overfit.py.",
deprecated_names=("LTX2_OVERFIT_DATA_DIR", ))
FASTVIDEO_TEST_LTX2_OVERFIT_CAPTION_JSON = EnvStr("videos2caption_1_sample.json",
category="test",
doc="Caption file, relative to the raw data directory.",
deprecated_names=("LTX2_OVERFIT_CAPTION_JSON", ))
FASTVIDEO_TEST_LTX2_OVERFIT_VIDEO_SUBDIR = EnvStr("video",
category="test",
doc="Video subdirectory, relative to the raw data directory.",
deprecated_names=("LTX2_OVERFIT_VIDEO_SUBDIR", ))
FASTVIDEO_TEST_LTX2_OVERFIT_OUTPUT_DIR = EnvStr("data/ltx2_overfit_preprocessed",
category="test",
doc="Output directory for preprocess_ltx2_overfit.py.",
deprecated_names=("LTX2_OVERFIT_OUTPUT_DIR", ))
FASTVIDEO_TEST_LTX2_OVERFIT_MODEL = EnvStr("FastVideo/LTX2-Distilled-Diffusers",
category="test",
doc="Model repository whose encoders preprocess_ltx2_overfit.py uses.",
deprecated_names=("LTX2_OVERFIT_MODEL", ))
FASTVIDEO_TEST_LTX2_OVERFIT_NUM_COPIES = EnvInt(4,
category="test",
doc="Number of copies of the overfit sample in the parquet file.",
deprecated_names=("LTX2_OVERFIT_NUM_COPIES", ))
FASTVIDEO_TEST_KANDINSKY5_OVERFIT_DATA_DIR = EnvStr("data/kandinsky5_overfit",
category="test",
doc="Raw data directory for preprocess_kandinsky5_overfit.py.",
deprecated_names=("KANDINSKY5_OVERFIT_DATA_DIR", ))
FASTVIDEO_TEST_KANDINSKY5_OVERFIT_OUTPUT_DIR = EnvStr("data/kandinsky5_overfit_preprocessed",
category="test",
doc="Output directory for preprocess_kandinsky5_overfit.py.",
deprecated_names=("KANDINSKY5_OVERFIT_OUTPUT_DIR", ))
# Variables that FastVideo no longer reads. Setting one logs a warning; delete
# the entries in the next minor release.
DEPRECATED_VARIABLES = {
"FASTVIDEO_TARGET_DEVICE": "no code reads it",
"FASTVIDEO_USE_PRECOMPILED": "no code reads it",
"FASTVIDEO_RINGBUFFER_WARNING_INTERVAL": "no code reads it",
"FASTVIDEO_ENGINE_ITERATION_TIMEOUT_S": "no code reads it",
"FASTVIDEO_SERVER_DEV_MODE": "no code reads it",
"FASTVIDEO_TEST_DYNAMO_FULLGRAPH_CAPTURE": "no code reads it",
"FASTVIDEO_TRACE_FUNCTION": "no code reads it",
}
def warn_deprecated_variables() -> None:
"""Log a warning for each variable in DEPRECATED_VARIABLES that is set."""
for name, reason in DEPRECATED_VARIABLES.items():
if name in os.environ:
_warn_once(f"{name} is deprecated and has no effect ({reason}); it will be removed in the next minor "
"release.")
def _register_fields() -> dict[str, EnvField]:
fastvideo/eval/__init__.py
+2 -2
View File
@@ -1,3 +1,4 @@
import fastvideo.envs as envs
from fastvideo.eval.models import ensure_checkpoint, get_cache_dir
@@ -14,9 +15,8 @@ def _redirect_third_party_caches() -> None:
env vars at all; their callsites in metric.py files pass
``download_root=str(get_cache_dir() / "<library>")`` directly.
"""
import os
root = get_cache_dir()
os.environ.setdefault("TORCH_HOME", str(root / "torch"))
envs.setdefault_external("TORCH_HOME", str(root / "torch"))
_redirect_third_party_caches()
fastvideo/eval/datasets/physics_iq.py
+2 -3
View File
@@ -27,7 +27,6 @@ pre-downloaded copy) to opt out of network fetches.
from __future__ import annotations
import csv
import os
from dataclasses import dataclass
from pathlib import Path
from urllib.request import urlretrieve
@@ -35,6 +34,7 @@ from urllib.request import urlretrieve
import cv2
import numpy as np
import fastvideo.envs as envs
from fastvideo.eval.datasets.base import PromptDataset
from fastvideo.eval.datasets.registry import register_dataset
from fastvideo.eval.models import get_cache_dir
@@ -59,11 +59,10 @@ _VENDORED_DESCRIPTIONS_CSV = (Path(__file__).resolve().parent.parent / "metrics"
# Public DeepMind bucket; HTTPS-readable, no auth. Override via
# ``FASTVIDEO_PHYSICS_IQ_BUCKET_URL`` (e.g. for an internal mirror).
_DEFAULT_BUCKET_URL = "https://storage.googleapis.com/physics-iq-benchmark"
def _bucket_url() -> str:
return os.environ.get("FASTVIDEO_PHYSICS_IQ_BUCKET_URL", _DEFAULT_BUCKET_URL)
return envs.FASTVIDEO_PHYSICS_IQ_BUCKET_URL.get()
def _default_dataset_root() -> Path:
fastvideo/eval/datasets/vbench.py
+4 -4
View File
@@ -7,9 +7,9 @@ keyed by dimension).
from __future__ import annotations
import json
import os
from pathlib import Path
import fastvideo.envs as envs
from fastvideo.eval.datasets.base import PromptDataset
from fastvideo.eval.datasets.registry import register_dataset
@@ -23,12 +23,12 @@ _FULL_INFO_REL = "fastvideo/third_party/eval/vbench/vbench/VBench_full_info.json
def _locate_full_info() -> Path:
env = os.environ.get("VBENCH_FULL_INFO_JSON")
env = envs.FASTVIDEO_VBENCH_FULL_INFO_JSON.get()
if env:
p = Path(env)
if p.is_file():
return p
raise FileNotFoundError(f"VBENCH_FULL_INFO_JSON={env} does not point at a file")
raise FileNotFoundError(f"FASTVIDEO_VBENCH_FULL_INFO_JSON={env} does not point at a file")
here = Path(__file__).resolve()
for ancestor in here.parents:
candidate = ancestor / _FULL_INFO_REL
@@ -38,7 +38,7 @@ def _locate_full_info() -> Path:
break
raise FileNotFoundError("Could not locate VBench_full_info.json. Initialize the upstream "
"submodule (`git submodule update --init "
"fastvideo/third_party/eval/vbench`) or set VBENCH_FULL_INFO_JSON.")
"fastvideo/third_party/eval/vbench`) or set FASTVIDEO_VBENCH_FULL_INFO_JSON.")
@register_dataset("vbench")
fastvideo/eval/metrics/audio/frechet_distance/metric.py
+2 -2
View File
@@ -11,7 +11,6 @@ two finite embeddings.
from __future__ import annotations
import os
from pathlib import Path
from typing import Any
@@ -19,6 +18,7 @@ import numpy as np
import torch
from scipy import linalg
import fastvideo.envs as envs
from fastvideo.eval.metrics.base import BaseMetric
from fastvideo.eval.registry import register
from fastvideo.eval.types import MetricResult
@@ -94,7 +94,7 @@ class FrechetAudioDistanceMetric(BaseMetric):
self._model: Any = None
self._gen_buf: list[np.ndarray] = []
self._ref_buf: list[np.ndarray] = []
self._cached_ref_path: str | None = os.environ.get(REF_FEATURES_ENV)
self._cached_ref_path: str | None = envs.FASTVIDEO_FAD_REF_FEATURES.get()
self._cached_ref_mu: np.ndarray | None = None
self._cached_ref_sigma: np.ndarray | None = None
self._n_cached_ref: int = 0
fastvideo/eval/metrics/common/fvd/metric.py
+2 -2
View File
@@ -75,13 +75,13 @@ import numpy as np
import scipy.linalg
import torch
import fastvideo.envs as envs
from fastvideo.eval.metrics.base import BaseMetric
from fastvideo.eval.metrics.common.fvd.extractors import (_BaseExtractor, available_extractors, load_extractor)
from fastvideo.eval.registry import register
from fastvideo.eval.types import MetricResult
_MIN_VIDEOS_WARN = 256 # below this FVD is unreliable
_REF_FEATURES_ENV = "FASTVIDEO_FVD_REF_FEATURES"
def _gaussian_params(features: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
@@ -220,7 +220,7 @@ class FVDMetric(BaseMetric):
# import. Precedence: kwarg > env-var > default.
if self._cache_path_arg is not None:
self.cache_path = os.path.expanduser(self._cache_path_arg)
elif env_path := os.environ.get(_REF_FEATURES_ENV):
elif env_path := envs.FASTVIDEO_FVD_REF_FEATURES.get():
self.cache_path = os.path.expanduser(env_path)
else:
self.cache_path = _default_cache_path(self._extractor_name)
fastvideo/eval/metrics/judge/third_person_separation/metric.py
+3 -4
View File
@@ -110,10 +110,9 @@ def _path_of(sample: dict, key: str) -> str | None:
def _resolve_api_key() -> str:
for env in ("GEMINI_API_KEY", "GOOGLE_API_KEY"):
key = os.environ.get(env)
if key:
return key.strip()
key = os.environ.get("GEMINI_API_KEY") or os.environ.get("GOOGLE_API_KEY")
if key:
return key.strip()
token = Path("~/.gemini_token").expanduser()
if token.is_file():
return token.read_text().strip()
fastvideo/eval/metrics/vbench/scene/metric.py
+2 -1
View File
@@ -21,6 +21,7 @@ import tempfile
import torch
import torchvision.io
import fastvideo.envs as envs
from fastvideo.eval.metrics.base import BaseMetric
from fastvideo.eval.registry import register
from fastvideo.eval.types import MetricResult
@@ -58,7 +59,7 @@ class SceneMetric(BaseMetric):
from transformers import Qwen2_5OmniForConditionalGeneration, Qwen2_5OmniProcessor
# AVoCaDO uses Qwen2.5-Omni — large multimodal model
os.environ.setdefault("VIDEO_MAX_PIXELS", str(20070400)) # 512*28*28*50
envs.setdefault_external("VIDEO_MAX_PIXELS", str(20070400)) # 512*28*28*50
self._model = Qwen2_5OmniForConditionalGeneration.from_pretrained(
self._model_path,
fastvideo/eval/models.py
+1 -4
View File
@@ -44,10 +44,7 @@ def get_cache_dir() -> Path:
``cache_dir``, etc.), pass ``str(get_cache_dir() / "<library>")`` so
users get a single ``FASTVIDEO_EVAL_CACHE`` knob to redirect them all.
"""
return Path(os.environ.get(
"FASTVIDEO_EVAL_CACHE",
os.path.join(envs.FASTVIDEO_CACHE_ROOT.get(), "eval"),
))
return Path(envs.FASTVIDEO_EVAL_CACHE.get())
def ensure_checkpoint(
fastvideo/fastvideo_args.py
+2
View File
@@ -361,6 +361,8 @@ class FastVideoArgs:
if env_backend is not None and backend_name_to_enum(env_backend) is not None:
self.attention_backend = env_backend
self._fold_vae_parallel_env()
import fastvideo.envs as envs
envs.warn_deprecated_variables()
self.check_fastvideo_args()
def _fold_vae_parallel_env(self) -> None:
fastvideo/logger.py
+39 -39
View File
@@ -15,11 +15,6 @@ from typing import Any, cast
import fastvideo.envs as envs
FASTVIDEO_CONFIGURE_LOGGING = envs.FASTVIDEO_CONFIGURE_LOGGING.get()
FASTVIDEO_LOGGING_CONFIG_PATH = envs.FASTVIDEO_LOGGING_CONFIG_PATH.get()
FASTVIDEO_LOGGING_LEVEL = envs.FASTVIDEO_LOGGING_LEVEL.get()
FASTVIDEO_LOGGING_PREFIX = envs.FASTVIDEO_LOGGING_PREFIX.get()
RED = '\033[91m'
GREEN = '\033[92m'
RESET = '\033[0;0m'
@@ -27,40 +22,43 @@ RESET = '\033[0;0m'
_warned_local_main_process = False
_warned_main_process = False
_FORMAT = (f"{FASTVIDEO_LOGGING_PREFIX}%(levelname)s %(asctime)s.%(msecs)03d "
_FORMAT = ("%(levelname)s %(asctime)s.%(msecs)03d "
"[%(filename)s:%(lineno)d] %(message)s")
_DATE_FORMAT = "%m-%d %H:%M:%S"
DEFAULT_LOGGING_CONFIG = {
"formatters": {
"fastvideo": {
"class": "fastvideo.logging_utils.NewLineFormatter",
"datefmt": _DATE_FORMAT,
"format": _FORMAT,
def _default_logging_config() -> dict[str, Any]:
"""Build the default logging configuration from FASTVIDEO_LOGGING_PREFIX and FASTVIDEO_LOGGING_LEVEL."""
return {
"formatters": {
"fastvideo": {
"class": "fastvideo.logging_utils.NewLineFormatter",
"datefmt": _DATE_FORMAT,
"format": envs.FASTVIDEO_LOGGING_PREFIX.get() + _FORMAT,
},
},
},
"handlers": {
"fastvideo": {
"class": "logging.StreamHandler",
"formatter": "fastvideo",
"level": FASTVIDEO_LOGGING_LEVEL,
"stream": "ext://sys.stdout",
"handlers": {
"fastvideo": {
"class": "logging.StreamHandler",
"formatter": "fastvideo",
"level": envs.FASTVIDEO_LOGGING_LEVEL.get(),
"stream": "ext://sys.stdout",
},
},
},
"loggers": {
"fastvideo": {
"loggers": {
"fastvideo": {
"handlers": ["fastvideo"],
"level": "DEBUG",
"propagate": False,
},
},
"root": {
"handlers": ["fastvideo"],
"level": "DEBUG",
"propagate": False,
},
},
"root": {
"handlers": ["fastvideo"],
"level": "DEBUG",
},
"version": 1,
"disable_existing_loggers": False
}
"version": 1,
"disable_existing_loggers": False
}
@lru_cache
@@ -180,20 +178,22 @@ class _FastvideoLogger(Logger):
def _configure_fastvideo_root_logger() -> None:
logging_config = dict[str, Any]()
configure_logging = envs.FASTVIDEO_CONFIGURE_LOGGING.get()
logging_config_path = envs.FASTVIDEO_LOGGING_CONFIG_PATH.get()
if not FASTVIDEO_CONFIGURE_LOGGING and FASTVIDEO_LOGGING_CONFIG_PATH:
if not configure_logging and logging_config_path:
raise RuntimeError("FASTVIDEO_CONFIGURE_LOGGING evaluated to false, but "
"FASTVIDEO_LOGGING_CONFIG_PATH was given. FASTVIDEO_LOGGING_CONFIG_PATH "
"implies FASTVIDEO_CONFIGURE_LOGGING. Please enable "
"FASTVIDEO_CONFIGURE_LOGGING or unset FASTVIDEO_LOGGING_CONFIG_PATH.")
if FASTVIDEO_CONFIGURE_LOGGING:
logging_config = DEFAULT_LOGGING_CONFIG
if configure_logging:
logging_config = _default_logging_config()
if FASTVIDEO_LOGGING_CONFIG_PATH:
if not path.exists(FASTVIDEO_LOGGING_CONFIG_PATH):
raise RuntimeError("Could not load logging config. File does not exist: %s", FASTVIDEO_LOGGING_CONFIG_PATH)
with open(FASTVIDEO_LOGGING_CONFIG_PATH, encoding="utf-8") as file:
if logging_config_path:
if not path.exists(logging_config_path):
raise RuntimeError("Could not load logging config. File does not exist: %s", logging_config_path)
with open(logging_config_path, encoding="utf-8") as file:
custom_config = json.loads(file.read())
if not isinstance(custom_config, dict):
@@ -286,7 +286,7 @@ def enable_trace_function_call(log_file_path: str, root_dir: str | None = None):
Note that this call is thread-level, any threads calling this function
will have the trace enabled. Other threads will not be affected.
"""
logger.warning("FASTVIDEO_TRACE_FUNCTION is enabled. It will record every"
logger.warning("Function-call tracing is enabled. It will record every"
" function executed by Python. This will slow down the code. It "
"is suggested to be used for debugging hang or crashes only.")
logger.info("Trace frame log is saved to %s", log_file_path)
fastvideo/mlx_runtime/fastwan.py
+6 -9
View File
@@ -6,7 +6,6 @@ from __future__ import annotations
import json
import math
import os
import statistics
import time
from dataclasses import dataclass
@@ -14,6 +13,7 @@ from pathlib import Path
from typing import TYPE_CHECKING, Any
from collections.abc import Callable
import fastvideo.envs as envs
from fastvideo.logger import init_logger
if TYPE_CHECKING:
@@ -326,7 +326,7 @@ def dq_gemm_engaged() -> int:
def affine_dq_gemm_min_m() -> int | None:
raw = os.environ.get("FASTVIDEO_MLX_DQ_GEMM", "1").strip().lower()
raw = envs.FASTVIDEO_MLX_DQ_GEMM.get().strip().lower()
if raw in {"", "0", "off", "false", "no"}:
return None
if raw in {"1", "on", "true", "yes"}:
@@ -400,9 +400,7 @@ def _use_fast_norm() -> bool:
traffic) and benchmark the speedup.
H3 uses its own fused RMSNorm default and does not consult this toggle.
"""
import os
return os.environ.get("FASTVIDEO_MLX_FAST_NORM", "0") == "1"
return envs.FASTVIDEO_MLX_FAST_NORM.get()
def layer_norm(x, weight=None, bias=None, eps: float = 1e-6):
@@ -633,12 +631,12 @@ class MLXWanTransformerBlock:
k_bh = key.transpose(0, 2, 1, 3)
v_bh = value.transpose(0, 2, 1, 3)
scale = self.head_dim**-0.5
window = int(os.environ.get("FASTVIDEO_MLX_WINDOW", "0") or "0")
window = envs.FASTVIDEO_MLX_WINDOW.get()
if window > 0:
from fastvideo.mlx_runtime.windowed_attention import windowed_attention
_warn_windowed_attention_once(window)
sink = int(os.environ.get("FASTVIDEO_MLX_WINDOW_SINK", "0") or "0")
sink = envs.FASTVIDEO_MLX_WINDOW_SINK.get()
attn_output = windowed_attention(q_bh, k_bh, v_bh, window=window, sink=sink, scale=scale)
else:
attn_output = mx.fast.scaled_dot_product_attention(q_bh, k_bh, v_bh, scale=scale)
@@ -693,7 +691,6 @@ class MLXWanDiT:
*,
compile: bool = False,
) -> None:
import os
self.weights = weights
self.blocks = blocks
@@ -711,7 +708,7 @@ class MLXWanDiT:
# good mx.compile target. Off by default so the eager path stays the
# baseline; enable via constructor or FASTVIDEO_MLX_COMPILE=1 and verify
# with the benchmark's SSIM ~= 1.0 check.
self._enable_compile = compile or os.environ.get("FASTVIDEO_MLX_COMPILE", "0") == "1"
self._enable_compile = compile or envs.FASTVIDEO_MLX_COMPILE.get()
self._compiled_forward: Callable[..., Any] | None = None
self._compiled_signature: tuple | None = None
fastvideo/mlx_runtime/wan22.py
+2 -2
View File
@@ -23,6 +23,7 @@ from pathlib import Path
from typing import TYPE_CHECKING, Any
from collections.abc import Callable
import fastvideo.envs as envs
from fastvideo.logger import init_logger
from fastvideo.mlx_runtime.fastwan import (
MLXWanT2VCrossAttention,
@@ -132,7 +133,6 @@ class MLXWan22DiT:
*,
compile: bool = False,
) -> None:
import os
self.weights = weights
self.blocks = blocks
@@ -144,7 +144,7 @@ class MLXWan22DiT:
self.patch_size = tuple(config["patch_size"])
self.out_channels = int(config["out_channels"])
self.eps = float(config.get("eps", 1e-6))
self._enable_compile = compile or os.environ.get("FASTVIDEO_MLX_COMPILE", "0") == "1"
self._enable_compile = compile or envs.FASTVIDEO_MLX_COMPILE.get()
self._compiled_forward: Callable[..., Any] | None = None
self._compiled_signature: tuple | None = None
fastvideo/models/encoders/gemma.py
+3 -2
View File
@@ -9,6 +9,7 @@ import torch
from torch import nn
from transformers import AutoTokenizer, Gemma3ForConditionalGeneration
import fastvideo.envs as envs
from fastvideo.configs.models.encoders import BaseEncoderOutput, TextEncoderConfig
from fastvideo.models.encoders.base import TextEncoder
from fastvideo.models.dits.ltx2 import (
@@ -27,7 +28,7 @@ import math
def _debug_gemma_log_line(message: str) -> None:
log_path = os.getenv("LTX2_FASTVIDEO_GEMMA_LOG", "")
log_path = envs.FASTVIDEO_LTX2_GEMMA_LOG.get()
if not log_path:
return
log_dir = os.path.dirname(log_path)
@@ -459,7 +460,7 @@ class LTX2GemmaTextEncoderModel(TextEncoder):
padding_side: str,
) -> tuple[torch.Tensor, torch.Tensor]:
encoded_text_features = torch.stack(hidden_states, dim=-1)
if os.getenv("LTX2_FASTVIDEO_GEMMA_LOG", ""):
if envs.FASTVIDEO_LTX2_GEMMA_LOG.get():
for idx, layer in enumerate(hidden_states):
_debug_gemma_log_line(f"fastvideo:gemma_hidden_state_{idx}"
f":sum={layer.float().sum().item():.6f}")
fastvideo/models/encoders/reason1.py
+2 -2
View File
@@ -1,13 +1,13 @@
# SPDX-License-Identifier: Apache-2.0
"""Reason1 (Qwen2.5-VL) text encoder."""
import os
from dataclasses import dataclass
from collections.abc import Iterable
import torch
from transformers import AutoProcessor
import fastvideo.envs as envs
from fastvideo.configs.models.encoders import BaseEncoderOutput, Reason1Config
from fastvideo.logger import init_logger
from fastvideo.models.encoders.base import TextEncoder
@@ -89,7 +89,7 @@ class Reason1TextEncoder(TextEncoder):
trust_remote_code=True,
)
weights_override = os.getenv("FASTVIDEO_REASON1_WEIGHTS_PATH")
weights_override = envs.FASTVIDEO_REASON1_WEIGHTS_PATH.get()
if weights_override:
self.secondary_weights = (_WeightsSource(
model_or_path=weights_override,
fastvideo/models/loader/benchmarks/benchmark_weight_loading.py
+5 -8
View File
@@ -21,6 +21,7 @@ import time
import torch
import torch.distributed as dist
import fastvideo.envs as envs
from fastvideo.distributed.parallel_state import (
cleanup_dist_env_and_memory,
get_node_group,
@@ -121,14 +122,10 @@ def main() -> None:
files = resolve_safetensors_files(load_path)
# Set default env vars for single-GPU standalone mode (no torchrun)
if "MASTER_ADDR" not in os.environ:
os.environ["MASTER_ADDR"] = "localhost"
if "MASTER_PORT" not in os.environ:
os.environ["MASTER_PORT"] = "29500"
if "RANK" not in os.environ:
os.environ["RANK"] = "0"
if "WORLD_SIZE" not in os.environ:
os.environ["WORLD_SIZE"] = "1"
envs.setdefault_external("MASTER_ADDR", "localhost")
envs.setdefault_external("MASTER_PORT", "29500")
envs.setdefault_external("RANK", "0")
envs.setdefault_external("WORLD_SIZE", "1")
maybe_init_distributed_environment_and_model_parallel(1, 1)
rank = dist.get_rank() if dist.is_initialized() else 0
fastvideo/models/loader/fsdp_load.py
+3 -4
View File
@@ -5,7 +5,6 @@
# Copyright 2025 The FastVideo Authors.
from __future__ import annotations
import os
import contextlib
import re
from collections.abc import Callable, Generator
@@ -392,7 +391,7 @@ def _regional_compile_unsupported_reason(
resolved = getattr(config, "_resolved_attention_backend", None)
resolved_name = getattr(resolved, "name", "")
if resolved_name == "VIDEO_SPARSE_ATTN_H3":
if os.environ.get("FASTVIDEO_H3_VSA_PROBE"):
if envs.FASTVIDEO_H3_VSA_PROBE.get():
return ("FASTVIDEO_H3_VSA_PROBE records tensors and files from the VSA-H3 attention body, which "
"regional fullgraph compile cannot capture; this model stays eager")
if not envs.FASTVIDEO_VSA_SM100A.get():
@@ -508,7 +507,7 @@ def shard_model(
ValueError: If no layer modules were sharded, indicating that no shard_condition was triggered.
"""
# Check if we should use size-based filtering
use_size_filtering = os.environ.get("FASTVIDEO_FSDP2_AUTOWRAP", "0") == "1"
use_size_filtering = envs.FASTVIDEO_FSDP2_AUTOWRAP.get()
if not fsdp_shard_conditions:
logger.warning("No FSDP shard conditions provided; nothing will be sharded.")
@@ -543,7 +542,7 @@ def shard_model(
if use_size_filtering:
# Size-based filtering mode
min_params = int(os.environ.get("FASTVIDEO_FSDP2_MIN_PARAMS", "10000000"))
min_params = envs.FASTVIDEO_FSDP2_MIN_PARAMS.get()
logger.info("Using size-based filtering with threshold: %.2fM", min_params / 1e6)
for n, m in reversed(named_modules):
fastvideo/models/vaes/ltx2vae.py
+2 -7
View File
@@ -7,7 +7,6 @@ from __future__ import annotations
import itertools
import math
import os
from dataclasses import dataclass, replace
from enum import Enum
from typing import Any, Callable, Iterator, List, NamedTuple, Tuple
@@ -17,14 +16,10 @@ import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange
import fastvideo.envs as envs
from fastvideo.models.vaes.common import DiagonalGaussianDistribution
def _is_env_enabled(name: str, default: str = "") -> bool:
value = os.getenv(name, default)
return value.lower() in {"1", "true", "yes", "on"}
# =============================================================================
# Enums
# =============================================================================
@@ -1595,7 +1590,7 @@ class LTX2CausalVideoAutoencoder(nn.Module):
self._use_tiling: bool = False
self._use_channels_last_3d: bool = False
if _is_env_enabled("FASTVIDEO_LTX2_VAE_CHANNELS_LAST_3D", default="1"):
if envs.FASTVIDEO_LTX2_VAE_CHANNELS_LAST_3D.get():
self.enable_channels_last_3d()
def _as_channels_last_3d(self, tensor: torch.Tensor) -> torch.Tensor:
fastvideo/pipelines/basic/ltx2/stages/ltx2_audio_decoding.py
+2 -3
View File
@@ -5,10 +5,9 @@ Audio decoding stage for LTX-2 pipelines.
from __future__ import annotations
import os
import torch
import fastvideo.envs as envs
from fastvideo.distributed import get_local_torch_device
from fastvideo.fastvideo_args import FastVideoArgs
from fastvideo.logger import init_logger
@@ -55,7 +54,7 @@ class LTX2AudioDecodingStage(PipelineStage):
self.vocoder = self.vocoder.to(device)
audio_latents = audio_latents.to(device)
disable_autocast = os.getenv("LTX2_DISABLE_AUDIO_AUTOCAST", "1") == "1"
disable_autocast = envs.FASTVIDEO_LTX2_DISABLE_AUDIO_AUTOCAST.get()
with torch.no_grad(), torch.autocast(
device_type="cuda",
dtype=audio_latents.dtype,
fastvideo/pipelines/basic/ltx2/stages/ltx2_denoising.py
+3 -3
View File
@@ -8,12 +8,12 @@ from __future__ import annotations
from contextlib import contextmanager
from itertools import combinations
import math
import os
from pathlib import Path
import torch
from tqdm.auto import tqdm
import fastvideo.envs as envs
from fastvideo.attention.backends.video_sparse_attn import (VideoSparseAttentionMetadataBuilder)
from fastvideo.attention.selector import component_attention_backend
from fastvideo.fastvideo_args import FastVideoArgs
@@ -54,7 +54,7 @@ except ImportError:
@contextmanager
def _nvtx_range(name: str):
if os.getenv("FASTVIDEO_NVTX_PROFILE", "0") == "1" and torch.cuda.is_available():
if envs.FASTVIDEO_NVTX_PROFILE.get() and torch.cuda.is_available():
torch.cuda.nvtx.range_push(name)
try:
yield
@@ -245,7 +245,7 @@ class LTX2DenoisingStage(PipelineStage):
else:
# Use distilled hardcoded schedule (or subsets) when enabled.
use_distilled_sigmas = (fastvideo_args.ltx2_use_distilled_sigmas
and os.getenv("LTX2_USE_DISTILLED_SIGMAS", "1") == "1")
and envs.FASTVIDEO_LTX2_USE_DISTILLED_SIGMAS.get())
max_distilled_steps = len(DISTILLED_SIGMA_VALUES) - 1
if use_distilled_sigmas and num_inference_steps <= max_distilled_steps:
sigmas, distilled_indices = _distilled_subset_sigmas(
fastvideo/pipelines/basic/magi_human/magi_human_pipeline.py
+2 -1
View File
@@ -29,6 +29,7 @@ from typing import Any
from transformers import AutoTokenizer
import fastvideo.envs as envs
from fastvideo.configs.models.encoders.t5gemma import T5GemmaEncoderConfig
from fastvideo.configs.models.vaes import OobleckVAEConfig
from fastvideo.fastvideo_args import FastVideoArgs
@@ -71,7 +72,7 @@ def _ensure_hf_token_env() -> str | None:
"""
token = resolve_hf_token()
if token:
os.environ.setdefault("HF_TOKEN", token)
envs.setdefault_external("HF_TOKEN", token)
return token
fastvideo/pipelines/preprocess/preprocess_kandinsky5_overfit.py
+16 -15
View File
@@ -26,7 +26,7 @@ Usage:
Input/output roots default to ``data/kandinsky5_overfit`` /
``data/kandinsky5_overfit_preprocessed`` and can be overridden with the
``KANDINSKY5_OVERFIT_DATA_DIR`` / ``KANDINSKY5_OVERFIT_OUTPUT_DIR`` env
``FASTVIDEO_TEST_KANDINSKY5_OVERFIT_DATA_DIR`` / ``FASTVIDEO_TEST_KANDINSKY5_OVERFIT_OUTPUT_DIR`` env
vars (used by the nightly e2e test to keep its disposable roots separate
from a user's real dataset at the defaults).
"""
@@ -34,9 +34,6 @@ from a user's real dataset at the defaults).
import json
import os
os.environ.setdefault("MASTER_ADDR", "localhost")
os.environ.setdefault("MASTER_PORT", "29520")
import cv2
import numpy as np
import pyarrow as pa
@@ -44,6 +41,7 @@ import pyarrow.parquet as pq
import torch
from transformers import AutoTokenizer
import fastvideo.envs as envs
from fastvideo.configs.pipelines.base import preprocess_text
from fastvideo.configs.pipelines.kandinsky5 import (
Kandinsky5T2VConfig,
@@ -70,11 +68,6 @@ MAX_WIDTH = 768
TRAIN_FPS = 24.0
MODEL_PATH = "kandinskylab/Kandinsky-5.0-T2V-Lite-sft-5s-Diffusers"
# Overridable so automation (e.g. the nightly e2e test) can point at its own
# test-owned directories instead of clobbering the documented default paths
# a user may have populated with their real dataset.
DATA_DIR = os.environ.get("KANDINSKY5_OVERFIT_DATA_DIR", "data/kandinsky5_overfit")
OUTPUT_DIR = os.environ.get("KANDINSKY5_OVERFIT_OUTPUT_DIR", "data/kandinsky5_overfit_preprocessed")
def load_video(path: str, num_frames: int, height: int, width: int) -> torch.Tensor:
@@ -143,17 +136,25 @@ def main() -> None:
# FastVideo's native text-encoder layers (e.g. CLIPAttention's
# QKVParallelLinear) are tensor-parallel-aware and assert the TP process
# group is initialized, even for a single-process/single-GPU run.
envs.setdefault_external("MASTER_ADDR", "localhost")
envs.setdefault_external("MASTER_PORT", "29520")
maybe_init_distributed_environment_and_model_parallel(1, 1)
# Overridable so automation (e.g. the nightly e2e test) can point at its own
# test-owned directories instead of clobbering the documented default paths
# a user may have populated with their real dataset.
data_dir = envs.FASTVIDEO_TEST_KANDINSKY5_OVERFIT_DATA_DIR.get()
output_dir = envs.FASTVIDEO_TEST_KANDINSKY5_OVERFIT_OUTPUT_DIR.get()
device = torch.device("cuda:0")
model_path = maybe_download_model(MODEL_PATH)
os.makedirs(OUTPUT_DIR, exist_ok=True)
os.makedirs(output_dir, exist_ok=True)
# Load captions. Validate up front (before spending minutes loading the
# VAE + two text encoders): an empty/malformed manifest would otherwise
# only surface as an IndexError on records[0] at parquet-write time.
manifest_path = os.path.join(DATA_DIR, "videos2caption.json")
manifest_path = os.path.join(data_dir, "videos2caption.json")
with open(manifest_path) as f:
caption_data = json.load(f)
if not isinstance(caption_data, list) or not caption_data:
@@ -212,7 +213,7 @@ def main() -> None:
for item in caption_data:
video_name = item["path"]
caption = get_caption(item)
video_path = os.path.join(DATA_DIR, "videos", video_name)
video_path = os.path.join(data_dir, "videos", video_name)
print(f"\nProcessing: {video_name}")
print(f" Caption: {caption[:80]}...")
@@ -289,7 +290,7 @@ def main() -> None:
for k in records[0]},
schema=pyarrow_schema_t2v,
)
output_path = os.path.join(OUTPUT_DIR, "data_00000.parquet")
output_path = os.path.join(output_dir, "data_00000.parquet")
pq.write_table(table, output_path)
print(f"\nWrote {len(records)} records to {output_path}")
@@ -297,12 +298,12 @@ def main() -> None:
# Wrap in "data" key -- ValidationDataset expects field="data".
# Use "caption" field -- ValidationDataset aliases it to "prompt".
val_prompts = {"data": [{"caption": get_caption(item)} for item in caption_data]}
val_path = os.path.join(OUTPUT_DIR, "validation_prompts.json")
val_path = os.path.join(output_dir, "validation_prompts.json")
with open(val_path, "w") as f:
json.dump(val_prompts, f, indent=2)
print(f"Wrote validation prompts to {val_path}")
print("\nDone! Use data_path: " + OUTPUT_DIR + " in training config.")
print("\nDone! Use data_path: " + output_dir + " in training config.")
if __name__ == "__main__":
fastvideo/pipelines/preprocess/preprocess_ltx2_overfit.py
+29 -28
View File
@@ -32,6 +32,7 @@ import pyarrow as pa
import pyarrow.parquet as pq
import torch
import fastvideo.envs as envs
from fastvideo.dataset.dataloader.schema import pyarrow_schema_t2v
from fastvideo.utils import maybe_download_model, verify_model_config_and_directory
@@ -41,25 +42,14 @@ MAX_HEIGHT = 480 # divisible by 32 (spatial compression)
MAX_WIDTH = 832
TRAIN_FPS = 24.0 # matches the LTX-2 preset fps used at validation
DATA_DIR = os.environ.get("LTX2_OVERFIT_DATA_DIR", "data/cats")
CAPTION_JSON = os.environ.get("LTX2_OVERFIT_CAPTION_JSON", "videos2caption_1_sample.json")
VIDEO_SUBDIR = os.environ.get("LTX2_OVERFIT_VIDEO_SUBDIR", "video")
OUTPUT_DIR = os.environ.get("LTX2_OVERFIT_OUTPUT_DIR", "data/ltx2_overfit_preprocessed")
MODEL_REPO = os.environ.get("LTX2_OVERFIT_MODEL", "FastVideo/LTX2-Distilled-Diffusers")
# The train dataloader samples with drop_last=True across data-parallel
# groups, so the dataset must hold at least num_sp_groups * batch_size
# rows or every rank gets zero batches. Replicate the overfit sample so
# a 4-GPU FSDP run still sees one batch per rank.
NUM_COPIES = int(os.environ.get("LTX2_OVERFIT_NUM_COPIES", "4"))
def _init_single_process_distributed() -> None:
"""FastVideo component loaders expect an initialized distributed env."""
os.environ.setdefault("MASTER_ADDR", "127.0.0.1")
os.environ.setdefault("MASTER_PORT", "29511")
os.environ.setdefault("RANK", "0")
os.environ.setdefault("WORLD_SIZE", "1")
os.environ.setdefault("LOCAL_RANK", "0")
envs.setdefault_external("MASTER_ADDR", "127.0.0.1")
envs.setdefault_external("MASTER_PORT", "29511")
envs.setdefault_external("RANK", "0")
envs.setdefault_external("WORLD_SIZE", "1")
envs.setdefault_external("LOCAL_RANK", "0")
from fastvideo.distributed import (
maybe_init_distributed_environment_and_model_parallel, )
maybe_init_distributed_environment_and_model_parallel(1, 1)
@@ -106,16 +96,27 @@ def main() -> None:
PipelineComponentLoader, )
from fastvideo.pipelines.basic.ltx2.pipeline_configs import LTX2T2VConfig
data_dir = envs.FASTVIDEO_TEST_LTX2_OVERFIT_DATA_DIR.get()
caption_json = envs.FASTVIDEO_TEST_LTX2_OVERFIT_CAPTION_JSON.get()
video_subdir = envs.FASTVIDEO_TEST_LTX2_OVERFIT_VIDEO_SUBDIR.get()
output_dir = envs.FASTVIDEO_TEST_LTX2_OVERFIT_OUTPUT_DIR.get()
model_repo = envs.FASTVIDEO_TEST_LTX2_OVERFIT_MODEL.get()
# The train dataloader samples with drop_last=True across data-parallel
# groups, so the dataset must hold at least num_sp_groups * batch_size
# rows or every rank gets zero batches. Replicate the overfit sample so
# a 4-GPU FSDP run still sees one batch per rank.
num_copies = envs.FASTVIDEO_TEST_LTX2_OVERFIT_NUM_COPIES.get()
device = torch.device("cuda:0")
model_path = maybe_download_model(MODEL_REPO)
model_path = maybe_download_model(model_repo)
model_index = verify_model_config_and_directory(model_path)
os.makedirs(OUTPUT_DIR, exist_ok=True)
os.makedirs(output_dir, exist_ok=True)
# The map-style dataset caches parquet file metadata; a stale cache
# next to a regenerated parquet can crash or serve old rows.
shutil.rmtree(os.path.join(OUTPUT_DIR, "map_style_cache"), ignore_errors=True)
shutil.rmtree(os.path.join(output_dir, "map_style_cache"), ignore_errors=True)
with open(os.path.join(DATA_DIR, CAPTION_JSON)) as f:
with open(os.path.join(data_dir, caption_json)) as f:
caption_data = json.load(f)
pipeline_config = LTX2T2VConfig()
@@ -163,7 +164,7 @@ def main() -> None:
video_name = item["path"]
record_id = f"{idx:04d}_{video_name}"
caption = item["cap"][0] if isinstance(item["cap"], list) else item["cap"]
video_path = os.path.join(DATA_DIR, VIDEO_SUBDIR, video_name)
video_path = os.path.join(data_dir, video_subdir, video_name)
print(f"\nProcessing: {video_name}")
print(f" Caption: {caption[:80]}...")
@@ -213,7 +214,7 @@ def main() -> None:
# Save the preprocessed clip so overfit tests can compare
# validation output against the memorization target.
import imageio
ref_path = os.path.join(OUTPUT_DIR, f"training_sample_{idx}.mp4")
ref_path = os.path.join(output_dir, f"training_sample_{idx}.mp4")
with imageio.get_writer(ref_path, fps=TRAIN_FPS) as writer:
for frame in frames_np:
writer.append_data(frame)
@@ -223,9 +224,9 @@ def main() -> None:
del text_encoder, tokenizer, vae
torch.cuda.empty_cache()
# Write parquet (replicated NUM_COPIES times; see comment at top)
# Write parquet (replicated num_copies times; see the comment above num_copies)
replicated = []
for copy_idx in range(max(1, NUM_COPIES)):
for copy_idx in range(max(1, num_copies)):
for r in records:
row = dict(r)
row["id"] = f"{r['id']}_copy{copy_idx}"
@@ -235,10 +236,10 @@ def main() -> None:
for k in replicated[0]},
schema=pyarrow_schema_t2v,
)
output_path = os.path.join(OUTPUT_DIR, "data_00000.parquet")
output_path = os.path.join(output_dir, "data_00000.parquet")
pq.write_table(table, output_path)
print(f"\nWrote {len(replicated)} records "
f"({len(records)} unique x {max(1, NUM_COPIES)} copies) to {output_path}")
f"({len(records)} unique x {max(1, num_copies)} copies) to {output_path}")
# Write validation prompts for the validation callback
val_prompts = {
@@ -246,12 +247,12 @@ def main() -> None:
"caption": (item["cap"][0] if isinstance(item["cap"], list) else item["cap"]),
} for item in caption_data],
}
val_path = os.path.join(OUTPUT_DIR, "validation_prompts.json")
val_path = os.path.join(output_dir, "validation_prompts.json")
with open(val_path, "w") as f:
json.dump(val_prompts, f, indent=2)
print(f"Wrote validation prompts to {val_path}")
print("\nDone! Use data_path: " + OUTPUT_DIR + " in training config.")
print("\nDone! Use data_path: " + output_dir + " in training config.")
if __name__ == "__main__":
fastvideo/pipelines/preprocess/preprocess_minimax_h3_overfit.py
+6 -6
View File
@@ -10,7 +10,6 @@ from __future__ import annotations
import argparse
import gc
import json
import os
from pathlib import Path
import shutil
from typing import Any
@@ -20,6 +19,7 @@ import pyarrow as pa
import pyarrow.parquet as pq
import torch
import fastvideo.envs as envs
from fastvideo.configs.pipelines.minimax_h3 import MiniMaxH3PipelineConfig
from fastvideo.dataset.dataloader.schema import pyarrow_schema_t2va
from fastvideo.fastvideo_args import FastVideoArgs
@@ -50,11 +50,11 @@ TRAINING_VIDEO_NAME = "1gGQy4nxyUo-Scene-016.mp4"
def _init_single_process_distributed() -> None:
"""Initialize the one-rank process groups required by component loaders."""
os.environ.setdefault("MASTER_ADDR", "127.0.0.1")
os.environ.setdefault("MASTER_PORT", "29531")
os.environ.setdefault("RANK", "0")
os.environ.setdefault("WORLD_SIZE", "1")
os.environ.setdefault("LOCAL_RANK", "0")
envs.setdefault_external("MASTER_ADDR", "127.0.0.1")
envs.setdefault_external("MASTER_PORT", "29531")
envs.setdefault_external("RANK", "0")
envs.setdefault_external("WORLD_SIZE", "1")
envs.setdefault_external("LOCAL_RANK", "0")
from fastvideo.distributed import maybe_init_distributed_environment_and_model_parallel
maybe_init_distributed_environment_and_model_parallel(1, 1)
fastvideo/pipelines/stages/denoising.py
+1 -3
View File
@@ -4,7 +4,6 @@ Denoising stage for diffusion pipelines.
"""
import inspect
import os
import weakref
from collections.abc import Iterable
from dataclasses import dataclass
@@ -141,8 +140,7 @@ class DenoisingStage(PipelineStage):
# gates should graduate to arch-config declarations like the precision
# policies above.
_is_flux = (getattr(fastvideo_args.pipeline_config.dit_config, "prefix", "") == "Flux")
if _is_flux and os.getenv("FASTVIDEO_FLUX2_DISABLE_BF16_REDUCED_PRECISION_REDUCTION",
"").lower() in {"1", "true", "yes"}:
if _is_flux and envs.FASTVIDEO_FLUX2_DISABLE_BF16_REDUCED_PRECISION_REDUCTION.get():
# Gate 1: tighten bf16 matmul accumulation for the 4-step Klein model (opt-in via env var).
torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction = False
# Gate 2: Flux2 runs its bf16 transformer WITHOUT autocast — autocast perturbs long-sequence attention enough to break 4-step latent parity.
fastvideo/pipelines/stages/latent_preparation.py
+2 -2
View File
@@ -3,13 +3,13 @@
Latent preparation stage for diffusion pipelines.
"""
import os
from typing import Any
import numpy as np
import torch
from diffusers.utils.torch_utils import randn_tensor
import fastvideo.envs as envs
from fastvideo.distributed import get_local_torch_device
from fastvideo.fastvideo_args import FastVideoArgs
from fastvideo.image_processor import ImageProcessor
@@ -475,7 +475,7 @@ class Cosmos25LatentPreparationStage(CosmosLatentPreparationStage):
)
video = torch.cat([cond_video, padding], dim=2)
if os.environ.get("FASTVIDEO_COSMOS25_LOG_KNOBS", "0") in ("1", "true", "True"):
if envs.FASTVIDEO_COSMOS25_LOG_KNOBS.get():
logger.info(
"[Cosmos2.5 latent_prep] using conditioning input: image=%s video_latent=%s video_path=%s | "
"num_video_frames=%s num_cond_frames=%s num_frames=%s",
fastvideo/tests/contract/test_env_policy.py
+81 -125
View File
@@ -6,7 +6,9 @@ The test parses every Python file under fastvideo/ (except fastvideo/third_party
and the registry fastvideo/envs.py) with ``ast`` and reports code that:
- reads the environment directly for a name outside EXTERNAL_ALLOWLIST;
- writes the environment directly (os.environ, os.putenv, monkeypatch.setenv);
- writes the environment directly (os.environ, os.putenv, monkeypatch.setenv),
or through the envs.*_external helpers for a name outside
EXTERNAL_WRITE_ALLOWLIST;
- uses the whole environment (os.environ.copy(), dict(os.environ), patch.dict);
- uses a registry field without calling one of its methods (``envs.X == "a"``,
``getter = envs.X.get``);
@@ -61,8 +63,33 @@ EXTERNAL_ALLOWLIST = {
"PATH": "Executable search path.",
"HF_TOKEN": "Token variable that huggingface_hub reads.",
"HUGGING_FACE_HUB_TOKEN": "Token variable that huggingface_hub reads.",
"GEMINI_API_KEY": "Gemini API key for the judge metrics.",
"GOOGLE_API_KEY": "Google API key, which the Gemini SDK also accepts.",
"CEREBRAS_API_KEY": "Cerebras API key for the streaming prompt provider.",
"GROQ_API_KEY": "Groq API key for the streaming prompt provider.",
"RAY_USAGE_STATS_ENABLED": "Ray usage-statistics switch.",
}
# Variables that FastVideo sets for other tools. Code writes them only through
# envs.set_external, envs.setdefault_external, or envs.unset_external, with a
# literal name.
EXTERNAL_WRITE_ALLOWLIST = {
"RANK": "torch.distributed rendezvous for single-process runs and workers.",
"LOCAL_RANK": "torch.distributed rendezvous for single-process runs and workers.",
"WORLD_SIZE": "torch.distributed rendezvous for single-process runs and workers.",
"MASTER_ADDR": "torch.distributed rendezvous for single-process runs.",
"MASTER_PORT": "torch.distributed rendezvous for single-process runs.",
"CUDA_VISIBLE_DEVICES": "Pins a streaming worker process to its GPU.",
"TORCH_NCCL_AVOID_RECORD_STREAMS": "Avoids NCCL record-stream memory growth in workers.",
"NCCL_ASYNC_ERROR_HANDLING": "Unset in workers because the value that Ray sets breaks graph building.",
"TORCH_HOME": "Points torch.hub at the evaluation cache.",
"HF_TOKEN": "Passes the resolved Hugging Face token to huggingface_hub.",
"VIDEO_MAX_PIXELS": "Pixel budget that qwen_vl_utils reads for Qwen2.5-Omni.",
"RAY_USAGE_STATS_ENABLED": "Turns off Ray usage statistics unless the user turned them on.",
"CUTE_DSL_ENABLE_TVM_FFI": "CUTLASS DSL switch that the NVFP4 FlashAttention-4 path needs.",
}
EXTERNAL_WRITE_HELPERS = {"set_external", "setdefault_external", "unset_external"}
# Methods that registry fields expose; ``get`` and ``is_set`` count as reads.
REGISTRY_METHODS = {"get", "set", "override", "is_set", "clear"}
REGISTRY_READ_METHODS = {"get", "is_set"}
@@ -72,111 +99,17 @@ REGISTRY_READ_METHODS = {"get", "is_set"}
# its violations are fixed; never add one. Kinds are described in
# docs/contributing/env_vars.md.
KNOWN_VIOLATIONS: dict[str, int] = {
'fastvideo/attention/backends/flash_attn.py: read FASTVIDEO_NVFP4_FA4': 1,
'fastvideo/attention/backends/video_sparse_attn_h3_probe.py: read FASTVIDEO_H3_VSA_PROBE': 1,
'fastvideo/attention/layer.py: read FASTVIDEO_DISABLE_ATTENTION_COMPILE': 2,
'fastvideo/attention/selector.py: read <dynamic>': 1,
'fastvideo/attention/utils/flash_attn_default.py: import-time-read FASTVIDEO_FA4': 1,
'fastvideo/benchmarks/eval_metalfx_rife.py: write FASTVIDEO_MLX_COMPILE': 1,
'fastvideo/benchmarks/mlx_fastwan_bench.py: read FASTVIDEO_MLX_COMPILE': 1,
'fastvideo/benchmarks/mlx_fastwan_bench.py: read FASTVIDEO_MLX_FAST_NORM': 1,
'fastvideo/benchmarks/mlx_fastwan_bench.py: write FASTVIDEO_MLX_COMPILE': 1,
'fastvideo/distributed/device_communicators/cpu_communicator.py: read VLLM_DIST_IDENT': 1,
'fastvideo/entrypoints/cli/utils.py: whole-environ': 1,
'fastvideo/entrypoints/openai/api_server.py: write FASTVIDEO_STAGE_LOGGING': 1,
'fastvideo/entrypoints/streaming/prompt/providers/cerebras.py: read <dynamic>': 1,
'fastvideo/entrypoints/streaming/prompt/providers/groq.py: read <dynamic>': 1,
'fastvideo/entrypoints/streaming/worker.py: write CUDA_VISIBLE_DEVICES': 1,
'fastvideo/entrypoints/video_generator.py: read FASTVIDEO_FFMPEG_BIN': 1,
'fastvideo/entrypoints/video_generator.py: read FASTVIDEO_NVENC_BF': 1,
'fastvideo/entrypoints/video_generator.py: read FASTVIDEO_NVENC_PRESET': 1,
'fastvideo/entrypoints/video_generator.py: read FASTVIDEO_NVENC_QP': 1,
'fastvideo/entrypoints/video_generator.py: read FASTVIDEO_NVENC_RC': 1,
'fastvideo/entrypoints/video_generator.py: read FASTVIDEO_NVENC_TUNE': 1,
'fastvideo/entrypoints/video_generator.py: read FASTVIDEO_OUTPUT_PIX_FMT': 1,
'fastvideo/entrypoints/video_generator.py: read FASTVIDEO_VIDEO_CODEC': 1,
'fastvideo/entrypoints/video_generator.py: read FASTVIDEO_X264_PRESET': 1,
'fastvideo/entrypoints/video_generator.py: write CUTE_DSL_ENABLE_TVM_FFI': 1,
'fastvideo/entrypoints/video_generator.py: write FASTVIDEO_NVFP4_FA4': 1,
'fastvideo/envs.py: prefix CMAKE_BUILD_TYPE': 1,
'fastvideo/envs.py: prefix CUDA_VISIBLE_DEVICES': 1,
'fastvideo/envs.py: prefix HCCL_SO_PATH': 1,
'fastvideo/envs.py: prefix LD_LIBRARY_PATH': 1,
'fastvideo/envs.py: prefix LOCAL_RANK': 1,
'fastvideo/envs.py: prefix MAX_JOBS': 1,
'fastvideo/envs.py: prefix NVCC_THREADS': 1,
'fastvideo/envs.py: prefix VERBOSE': 1,
'fastvideo/envs.py: unread CMAKE_BUILD_TYPE': 1,
'fastvideo/envs.py: unread CUDA_VISIBLE_DEVICES': 1,
'fastvideo/envs.py: unread FASTVIDEO_ENGINE_ITERATION_TIMEOUT_S': 1,
'fastvideo/envs.py: unread FASTVIDEO_RINGBUFFER_WARNING_INTERVAL': 1,
'fastvideo/envs.py: unread FASTVIDEO_SERVER_DEV_MODE': 1,
'fastvideo/envs.py: unread FASTVIDEO_TARGET_DEVICE': 1,
'fastvideo/envs.py: unread FASTVIDEO_TEST_DYNAMO_FULLGRAPH_CAPTURE': 1,
'fastvideo/envs.py: unread FASTVIDEO_TRACE_FUNCTION': 1,
'fastvideo/envs.py: unread FASTVIDEO_USE_PRECOMPILED': 1,
'fastvideo/envs.py: unread LD_LIBRARY_PATH': 1,
'fastvideo/envs.py: unread MAX_JOBS': 1,
'fastvideo/envs.py: unread NVCC_THREADS': 1,
'fastvideo/envs.py: unread VERBOSE': 1,
'fastvideo/eval/__init__.py: write TORCH_HOME': 1,
'fastvideo/eval/datasets/physics_iq.py: read FASTVIDEO_PHYSICS_IQ_BUCKET_URL': 1,
'fastvideo/eval/datasets/vbench.py: read VBENCH_FULL_INFO_JSON': 1,
'fastvideo/eval/metrics/audio/frechet_distance/metric.py: read <dynamic>': 1,
'fastvideo/eval/metrics/common/fvd/metric.py: read <dynamic>': 1,
'fastvideo/eval/metrics/judge/third_person_separation/metric.py: read <dynamic>': 1,
'fastvideo/eval/metrics/vbench/scene/metric.py: write VIDEO_MAX_PIXELS': 1,
'fastvideo/eval/models.py: read FASTVIDEO_EVAL_CACHE': 1,
'fastvideo/logger.py: import-time-read FASTVIDEO_CONFIGURE_LOGGING': 1,
'fastvideo/logger.py: import-time-read FASTVIDEO_LOGGING_CONFIG_PATH': 1,
'fastvideo/logger.py: import-time-read FASTVIDEO_LOGGING_LEVEL': 1,
'fastvideo/logger.py: import-time-read FASTVIDEO_LOGGING_PREFIX': 1,
'fastvideo/mlx_runtime/fastwan.py: read FASTVIDEO_MLX_COMPILE': 1,
'fastvideo/mlx_runtime/fastwan.py: read FASTVIDEO_MLX_DQ_GEMM': 1,
'fastvideo/mlx_runtime/fastwan.py: read FASTVIDEO_MLX_FAST_NORM': 1,
'fastvideo/mlx_runtime/fastwan.py: read FASTVIDEO_MLX_WINDOW': 1,
'fastvideo/mlx_runtime/fastwan.py: read FASTVIDEO_MLX_WINDOW_SINK': 1,
'fastvideo/mlx_runtime/memory.py: write <dynamic>': 1,
'fastvideo/mlx_runtime/wan22.py: read FASTVIDEO_MLX_COMPILE': 1,
'fastvideo/models/encoders/gemma.py: read LTX2_FASTVIDEO_GEMMA_LOG': 2,
'fastvideo/models/encoders/reason1.py: read FASTVIDEO_REASON1_WEIGHTS_PATH': 1,
'fastvideo/models/loader/benchmarks/benchmark_weight_loading.py: write MASTER_ADDR': 1,
'fastvideo/models/loader/benchmarks/benchmark_weight_loading.py: write MASTER_PORT': 1,
'fastvideo/models/loader/benchmarks/benchmark_weight_loading.py: write RANK': 1,
'fastvideo/models/loader/benchmarks/benchmark_weight_loading.py: write WORLD_SIZE': 1,
'fastvideo/models/loader/fsdp_load.py: read FASTVIDEO_FSDP2_AUTOWRAP': 1,
'fastvideo/models/loader/fsdp_load.py: read FASTVIDEO_FSDP2_MIN_PARAMS': 1,
'fastvideo/models/loader/fsdp_load.py: read FASTVIDEO_H3_VSA_PROBE': 1,
'fastvideo/models/vaes/ltx2vae.py: read <dynamic>': 1,
'fastvideo/performance/hf_store.py: read HF_REPO_ID': 1,
'fastvideo/performance/hf_store.py: read PERFORMANCE_TRACKING_SYNC_REUSE_TTL_SECONDS': 1,
'fastvideo/performance_dashboard/api.py: read PERFORMANCE_TRACKING_ROOT': 1,
'fastvideo/pipelines/basic/ltx2/stages/ltx2_audio_decoding.py: read LTX2_DISABLE_AUDIO_AUTOCAST': 1,
'fastvideo/pipelines/basic/ltx2/stages/ltx2_denoising.py: read FASTVIDEO_NVTX_PROFILE': 1,
'fastvideo/pipelines/basic/ltx2/stages/ltx2_denoising.py: read LTX2_USE_DISTILLED_SIGMAS': 1,
'fastvideo/pipelines/basic/magi_human/magi_human_pipeline.py: write HF_TOKEN': 1,
'fastvideo/pipelines/preprocess/preprocess_kandinsky5_overfit.py: read KANDINSKY5_OVERFIT_DATA_DIR': 1,
'fastvideo/pipelines/preprocess/preprocess_kandinsky5_overfit.py: read KANDINSKY5_OVERFIT_OUTPUT_DIR': 1,
'fastvideo/pipelines/preprocess/preprocess_kandinsky5_overfit.py: write MASTER_ADDR': 1,
'fastvideo/pipelines/preprocess/preprocess_kandinsky5_overfit.py: write MASTER_PORT': 1,
'fastvideo/pipelines/preprocess/preprocess_ltx2_overfit.py: read LTX2_OVERFIT_CAPTION_JSON': 1,
'fastvideo/pipelines/preprocess/preprocess_ltx2_overfit.py: read LTX2_OVERFIT_DATA_DIR': 1,
'fastvideo/pipelines/preprocess/preprocess_ltx2_overfit.py: read LTX2_OVERFIT_MODEL': 1,
'fastvideo/pipelines/preprocess/preprocess_ltx2_overfit.py: read LTX2_OVERFIT_NUM_COPIES': 1,
'fastvideo/pipelines/preprocess/preprocess_ltx2_overfit.py: read LTX2_OVERFIT_OUTPUT_DIR': 1,
'fastvideo/pipelines/preprocess/preprocess_ltx2_overfit.py: read LTX2_OVERFIT_VIDEO_SUBDIR': 1,
'fastvideo/pipelines/preprocess/preprocess_ltx2_overfit.py: write LOCAL_RANK': 1,
'fastvideo/pipelines/preprocess/preprocess_ltx2_overfit.py: write MASTER_ADDR': 1,
'fastvideo/pipelines/preprocess/preprocess_ltx2_overfit.py: write MASTER_PORT': 1,
'fastvideo/pipelines/preprocess/preprocess_ltx2_overfit.py: write RANK': 1,
'fastvideo/pipelines/preprocess/preprocess_ltx2_overfit.py: write WORLD_SIZE': 1,
'fastvideo/pipelines/preprocess/preprocess_minimax_h3_overfit.py: write LOCAL_RANK': 1,
'fastvideo/pipelines/preprocess/preprocess_minimax_h3_overfit.py: write MASTER_ADDR': 1,
'fastvideo/pipelines/preprocess/preprocess_minimax_h3_overfit.py: write MASTER_PORT': 1,
'fastvideo/pipelines/preprocess/preprocess_minimax_h3_overfit.py: write RANK': 1,
'fastvideo/pipelines/preprocess/preprocess_minimax_h3_overfit.py: write WORLD_SIZE': 1,
'fastvideo/pipelines/stages/denoising.py: read FASTVIDEO_FLUX2_DISABLE_BF16_REDUCED_PRECISION_REDUCTION': 1,
'fastvideo/pipelines/stages/latent_preparation.py: read FASTVIDEO_COSMOS25_LOG_KNOBS': 1,
'fastvideo/tests/api/test_attention_selector_resolution.py: write FASTVIDEO_ATTENTION_BACKEND': 17,
'fastvideo/tests/attention/test_flash_attn_no_pad_resolver.py: write FASTVIDEO_FA4': 3,
'fastvideo/tests/attention/test_flash_attn_nvfp4_opt_out.py: write FASTVIDEO_NVFP4_FA4': 4,
@@ -345,8 +278,8 @@ KNOWN_VIOLATIONS: dict[str, int] = {
'fastvideo/tests/modal/test_pr_test.py: write BUILDKITE_REPO': 2,
'fastvideo/tests/nightly/test_e2e_dmd_t2v_crush_smol.py: write WANDB_MODE': 1,
'fastvideo/tests/nightly/test_e2e_i2v_overfit_single_sample.py: write WANDB_MODE': 1,
'fastvideo/tests/nightly/test_e2e_kandinsky5_dmd_t2v_overfit.py: read <dynamic>': 1,
'fastvideo/tests/nightly/test_e2e_kandinsky5_dmd_t2v_overfit.py: read KANDINSKY5_E2E_NUM_GPUS': 1,
'fastvideo/tests/nightly/test_e2e_kandinsky5_dmd_t2v_overfit.py: read KANDINSKY5_E2E_WRITE_REFERENCE': 1,
'fastvideo/tests/nightly/test_e2e_kandinsky5_dmd_t2v_overfit.py: whole-environ': 3,
'fastvideo/tests/nightly/test_e2e_kandinsky5_dmd_t2v_overfit.py: write WANDB_MODE': 1,
'fastvideo/tests/nightly/test_e2e_ltx2_overfit_new_stack.py: read FASTVIDEO_NIGHTLY': 1,
@@ -420,8 +353,10 @@ KNOWN_VIOLATIONS: dict[str, int] = {
'fastvideo/tests/ssim/conftest.py: write <dynamic>': 3,
'fastvideo/tests/ssim/inference_similarity_utils.py: read FASTVIDEO_ATTENTION_BACKEND': 1,
'fastvideo/tests/ssim/inference_similarity_utils.py: write FASTVIDEO_ATTENTION_BACKEND': 3,
'fastvideo/tests/ssim/reference_utils.py: read <dynamic>': 1,
'fastvideo/tests/ssim/reference_videos_cli.py: read <dynamic>': 3,
'fastvideo/tests/ssim/reference_utils.py: read FASTVIDEO_SSIM_FULL_QUALITY': 1,
'fastvideo/tests/ssim/reference_videos_cli.py: read <dynamic>': 1,
'fastvideo/tests/ssim/reference_videos_cli.py: read FASTVIDEO_SSIM_REFERENCE_HF_REPO': 1,
'fastvideo/tests/ssim/reference_videos_cli.py: read FASTVIDEO_SSIM_REFERENCE_HF_REPO_TYPE': 1,
'fastvideo/tests/ssim/test_dreamx_world_similarity.py: read DREAMX_WORLD_AR_SSIM_MODEL_PATH': 1,
'fastvideo/tests/ssim/test_dreamx_world_similarity.py: read DREAMX_WORLD_SSIM_MODEL_PATH': 1,
'fastvideo/tests/ssim/test_flux_t2i_similarity.py: read FLUX_T2I_MODEL_DIR': 1,
@@ -444,7 +379,7 @@ KNOWN_VIOLATIONS: dict[str, int] = {
'fastvideo/tests/stages/_denoising_fixtures.py: write FASTVIDEO_CFG_GATE_STEP': 2,
'fastvideo/tests/stages/test_kandinsky5_dmd_stage_backend_engages.py: write FASTVIDEO_ATTENTION_BACKEND': 1,
'fastvideo/tests/stages/test_minimax_h3_encoding_offload.py: write FASTVIDEO_ATTENTION_BACKEND': 1,
'fastvideo/tests/train/methods/grad_norm_regression.py: read <dynamic>': 1,
'fastvideo/tests/train/methods/grad_norm_regression.py: read FASTVIDEO_GRADNORM_UPDATE': 1,
'fastvideo/tests/train/methods/test_cosmos_finetune.py: write MASTER_ADDR': 1,
'fastvideo/tests/train/methods/test_cosmos_finetune.py: write MASTER_PORT': 1,
'fastvideo/tests/train/methods/test_ltx2_finetune.py: write MASTER_ADDR': 1,
@@ -534,23 +469,11 @@ KNOWN_VIOLATIONS: dict[str, int] = {
'fastvideo/tests/vaes/test_wan_vae.py: write MASTER_PORT': 1,
'fastvideo/tests/worker/test_gpu_worker.py: write FASTVIDEO_NVTX_PROFILE': 2,
'fastvideo/tests/worker/test_gpu_worker.py: write LOCAL_RANK': 1,
'fastvideo/train/entrypoint/dcp_to_diffusers.py: write <dynamic>': 1,
'fastvideo/train/entrypoint/train.py: write FASTVIDEO_ATTENTION_BACKEND': 2,
'fastvideo/training/self_forcing_distillation_pipeline.py: read FASTVIDEO_FSDP2_AUTOWRAP': 1,
'fastvideo/utils.py: read <dynamic>': 5,
'fastvideo/utils.py: read FASTVIDEO_WORKER_MULTIPROC_METHOD': 1,
'fastvideo/utils.py: write <dynamic>': 1,
'fastvideo/utils.py: write FASTVIDEO_WORKER_MULTIPROC_METHOD': 1,
'fastvideo/worker/gpu_worker.py: write LOCAL_RANK': 1,
'fastvideo/worker/gpu_worker.py: write NCCL_ASYNC_ERROR_HANDLING': 1,
'fastvideo/worker/gpu_worker.py: write RANK': 1,
'fastvideo/worker/gpu_worker.py: write TORCH_NCCL_AVOID_RECORD_STREAMS': 1,
'fastvideo/worker/gpu_worker.py: write WORLD_SIZE': 1,
'fastvideo/worker/ray_distributed_executor.py: read <dynamic>': 2,
'fastvideo/worker/ray_distributed_executor.py: read RAY_USAGE_STATS_ENABLED': 1,
'fastvideo/worker/ray_distributed_executor.py: whole-environ': 1,
'fastvideo/worker/ray_distributed_executor.py: write RAY_USAGE_STATS_ENABLED': 1,
'fastvideo/worker/ray_env.py: import-time-read FASTVIDEO_CONFIG_ROOT': 1,
'fastvideo/worker/ray_env.py: read <dynamic>': 1,
'fastvideo/worker/ray_utils.py: whole-environ': 1,
'fastvideo/worker/worker_base.py: read <dynamic>': 1,
@@ -576,12 +499,6 @@ def is_allowlisted(name: str) -> bool:
return False
def _literal(node: ast.AST | None) -> str:
if isinstance(node, ast.Constant) and isinstance(node.value, str):
return node.value
return "<dynamic>"
class EnvAccessScanner:
"""Find environment accesses and registry-field uses in one module.
@@ -604,6 +521,13 @@ class EnvAccessScanner:
self.getenv_names: set[str] = set()
self.setenv_names: set[str] = set()
self.envs_module_names: set[str] = set()
# Module-level string constants, so that os.environ.get(NAME_ENV) resolves to the name.
self.constants: dict[str, str] = {}
for statement in tree.body:
if (isinstance(statement, ast.Assign) and len(statement.targets) == 1
and isinstance(statement.targets[0], ast.Name) and isinstance(statement.value, ast.Constant)
and isinstance(statement.value.value, str)):
self.constants[statement.targets[0].id] = statement.value.value
for node in ast.walk(tree):
if isinstance(node, ast.Import):
for alias in node.names:
@@ -624,6 +548,14 @@ class EnvAccessScanner:
elif node.module == "fastvideo" and alias.name == "envs":
self.envs_module_names.add(local)
def _literal(self, node: ast.AST | None) -> str:
"""Return the variable name that a call argument spells, or "<dynamic>"."""
if isinstance(node, ast.Constant) and isinstance(node.value, str):
return node.value
if isinstance(node, ast.Name) and node.id in self.constants:
return self.constants[node.id]
return "<dynamic>"
def _is_os_attr(self, node: ast.AST, attrs: tuple[str, ...]) -> bool:
return (isinstance(node, ast.Attribute) and node.attr in attrs and isinstance(node.value, ast.Name)
and node.value.id in self.os_names)
@@ -670,24 +602,24 @@ class EnvAccessScanner:
call = self.parents[parent]
assert isinstance(call, ast.Call)
if call.func is parent and parent.attr == "get":
self.violations.append(("read", _literal(call.args[0] if call.args else None), line))
self.violations.append(("read", self._literal(call.args[0] if call.args else None), line))
return
if call.func is parent and parent.attr in ("setdefault", "pop"):
self.violations.append(("write", _literal(call.args[0] if call.args else None), line))
self.violations.append(("write", self._literal(call.args[0] if call.args else None), line))
return
if isinstance(parent, ast.Subscript) and parent.value is node:
kind = "read" if isinstance(parent.ctx, ast.Load) else "write"
self.violations.append((kind, _literal(parent.slice), line))
self.violations.append((kind, self._literal(parent.slice), line))
return
if (isinstance(parent, ast.Compare) and len(parent.ops) == 1 and isinstance(parent.ops[0], (ast.In, ast.NotIn))
and parent.comparators[0] is node):
self.violations.append(("read", _literal(parent.left), line))
self.violations.append(("read", self._literal(parent.left), line))
return
self.violations.append(("whole-environ", "", line))
def _check_call(self, node: ast.Call) -> None:
func = node.func
name = _literal(node.args[0] if node.args else None)
name = self._literal(node.args[0] if node.args else None)
if self._is_os_attr(func, ("getenv", "getenvb")) or (isinstance(func, ast.Name)
and func.id in self.getenv_names):
self.violations.append(("read", name, node.lineno))
@@ -695,6 +627,9 @@ class EnvAccessScanner:
or (isinstance(func, ast.Name) and func.id in self.setenv_names)
or (isinstance(func, ast.Attribute) and func.attr in ("setenv", "delenv"))):
self.violations.append(("write", name, node.lineno))
elif (isinstance(func, ast.Attribute) and func.attr in EXTERNAL_WRITE_HELPERS
and self._is_envs_module(func.value) and name not in EXTERNAL_WRITE_ALLOWLIST):
self.violations.append(("write", name, node.lineno))
def _check_registry_use(self, node: ast.Attribute, parent: ast.AST | None, in_function: bool) -> None:
# Only a call counts: ``getter = envs.X.get`` neither reads the variable
@@ -715,11 +650,21 @@ def scanned_files() -> list[Path]:
if path != REGISTRY_PATH and not any(excluded in path.parents for excluded in EXCLUDED_DIRS))
def registry_internal_reads(registry_names: set[str]) -> Counter[str]:
"""Count registry reads inside fastvideo/envs.py itself, such as a computed default that reads another field."""
reads: Counter[str] = Counter()
for node in ast.walk(ast.parse(REGISTRY_PATH.read_text(encoding="utf-8"))):
if (isinstance(node, ast.Attribute) and node.attr in REGISTRY_READ_METHODS and isinstance(node.value, ast.Name)
and node.value.id in registry_names):
reads[node.value.id] += 1
return reads
def collect_violations(registry: ModuleType) -> tuple[dict[str, list[int]], Counter[str]]:
"""Return every violation key with its line numbers, plus registry read counts."""
registry_names = set(registry.environment_variables)
found: dict[str, list[int]] = {}
reads: Counter[str] = Counter()
reads: Counter[str] = registry_internal_reads(registry_names)
for path in scanned_files():
relative = path.relative_to(REPO_ROOT).as_posix()
scanner = EnvAccessScanner(ast.parse(path.read_text(encoding="utf-8"), filename=relative), registry_names)
@@ -793,10 +738,21 @@ def _escape(text: str) -> str:
def render_env_table(registry: ModuleType) -> str:
"""Render the registry as an aligned Markdown table, in declaration order."""
"""Render the registry and the deprecation table as aligned Markdown tables, in declaration order."""
rows = [["Variable", "Type", "Default", "Category", "Description"]]
for name, field in registry.environment_variables.items():
rows.append([f"`{name}`", field.type_name, _render_default(field), field.category, _escape(field.doc)])
doc = _escape(field.doc)
if field.deprecated_names:
doc += " Deprecated names: " + ", ".join(f"`{old}`" for old in field.deprecated_names) + "."
rows.append([f"`{name}`", field.type_name, _render_default(field), field.category, doc])
deprecated_rows = [["Deprecated variable", "Reason"]]
deprecated_rows.extend([f"`{name}`", reason] for name, reason in registry.DEPRECATED_VARIABLES.items())
return (_render_table(rows) + "\n\nVariables that FastVideo no longer reads; setting one logs a warning:\n\n" +
_render_table(deprecated_rows))
def _render_table(rows: list[list[str]]) -> str:
"""Render rows as a Markdown table whose pipes line up."""
widths = [max(len(row[column]) for row in rows) for column in range(len(rows[0]))]
def render_row(cells: list[str]) -> str:
fastvideo/tests/nightly/test_e2e_kandinsky5_dmd_t2v_overfit.py
+3 -3
View File
@@ -84,7 +84,7 @@ NUM_GPUS = os.environ.get("KANDINSKY5_E2E_NUM_GPUS", "1")
# with their real videos/captions for the actual recipe:
# _clean_previous_artifacts() deletes this root wholesale on every run,
# and the preprocess subprocess is pointed here via the
# KANDINSKY5_OVERFIT_DATA_DIR / KANDINSKY5_OVERFIT_OUTPUT_DIR env
# FASTVIDEO_TEST_KANDINSKY5_OVERFIT_DATA_DIR / FASTVIDEO_TEST_KANDINSKY5_OVERFIT_OUTPUT_DIR env
# overrides instead of the script's user-facing defaults.
E2E_ROOT = Path("data") / "kandinsky5_e2e"
RAW_DATA_DIR = E2E_ROOT / "raw"
@@ -223,8 +223,8 @@ def _run_preprocessing() -> None:
# (data/kandinsky5_overfit{,_preprocessed}) -- the very directories a
# user populates for the real recipe, which this test must never touch.
env = dict(os.environ)
env["KANDINSKY5_OVERFIT_DATA_DIR"] = str(RAW_DATA_DIR)
env["KANDINSKY5_OVERFIT_OUTPUT_DIR"] = str(PREPROCESSED_DIR)
env["FASTVIDEO_TEST_KANDINSKY5_OVERFIT_DATA_DIR"] = str(RAW_DATA_DIR)
env["FASTVIDEO_TEST_KANDINSKY5_OVERFIT_OUTPUT_DIR"] = str(PREPROCESSED_DIR)
cmd = [
sys.executable,
"-m",
fastvideo/tests/nightly/test_e2e_ltx2_overfit_new_stack.py
+3 -3
View File
@@ -82,9 +82,9 @@ def download_data():
def run_preprocessing(case: dict):
env = os.environ.copy()
env["CUDA_VISIBLE_DEVICES"] = env.get("CUDA_VISIBLE_DEVICES", "0").split(",")[0]
env["LTX2_OVERFIT_DATA_DIR"] = str(LOCAL_RAW_DATA_DIR)
env["LTX2_OVERFIT_OUTPUT_DIR"] = str(case["prep_dir"])
env["LTX2_OVERFIT_MODEL"] = case["model"]
env["FASTVIDEO_TEST_LTX2_OVERFIT_DATA_DIR"] = str(LOCAL_RAW_DATA_DIR)
env["FASTVIDEO_TEST_LTX2_OVERFIT_OUTPUT_DIR"] = str(case["prep_dir"])
env["FASTVIDEO_TEST_LTX2_OVERFIT_MODEL"] = case["model"]
cmd = [sys.executable, PREPROCESSING_SCRIPT]
print(f"Running preprocessing: {cmd} (model={case['model']})")
subprocess.run(cmd, check=True, env=env)
fastvideo/train/entrypoint/dcp_to_diffusers.py
+6 -8
View File
@@ -33,6 +33,7 @@ import os
import sys
from typing import Any
import fastvideo.envs as envs
from fastvideo.logger import init_logger
logger = init_logger(__name__)
@@ -45,14 +46,11 @@ def _ensure_distributed() -> None:
For plain ``python`` we fill in the minimum required vars so
that ``init_process_group`` succeeds with world_size=1.
"""
for key, default in [
("RANK", "0"),
("LOCAL_RANK", "0"),
("WORLD_SIZE", "1"),
("MASTER_ADDR", "127.0.0.1"),
("MASTER_PORT", "29500"),
]:
os.environ.setdefault(key, default)
envs.setdefault_external("RANK", "0")
envs.setdefault_external("LOCAL_RANK", "0")
envs.setdefault_external("WORLD_SIZE", "1")
envs.setdefault_external("MASTER_ADDR", "127.0.0.1")
envs.setdefault_external("MASTER_PORT", "29500")
def _save_role_pretrained(
fastvideo/train/models/ltx2/ltx2.py
+1 -1
View File
@@ -396,7 +396,7 @@ class LTX2Model(WanModel):
f"checkpoint's expected text context dim {expected_dim}. "
"The parquet was likely preprocessed with a different "
"LTX-2 version (2.0 stores 3840-d, 2.3 stores 4096-d); "
"re-run preprocess_ltx2_overfit.py with LTX2_OVERFIT_MODEL "
"re-run preprocess_ltx2_overfit.py with FASTVIDEO_TEST_LTX2_OVERFIT_MODEL "
"set to the same checkpoint as models.student.init_from.")
def _update_rope_fps(self, infos: list[dict[str, Any]] | None) -> None:
fastvideo/training/self_forcing_distillation_pipeline.py
+1 -2
View File
@@ -1,6 +1,5 @@
# SPDX-License-Identifier: Apache-2.0
import copy
import os
import time
from collections import deque
from typing import Any
@@ -47,7 +46,7 @@ class SelfForcingDistillationPipeline(DistillationPipeline):
def initialize_training_pipeline(self, training_args: TrainingArgs):
"""Initialize the self-forcing training pipeline."""
# Check if FSDP2 auto wrap is enabled - not supported for self-forcing distillation
if os.environ.get("FASTVIDEO_FSDP2_AUTOWRAP", "0") == "1":
if envs.FASTVIDEO_FSDP2_AUTOWRAP.get():
raise NotImplementedError("FASTVIDEO_FSDP2_AUTOWRAP is not implemented for self-forcing distillation. "
"Please set FASTVIDEO_FSDP2_AUTOWRAP=0 or unset the environment variable.")
fastvideo/utils.py
+5 -5
View File
@@ -82,16 +82,16 @@ def find_nccl_library() -> str:
def find_hccl_library() -> str:
"""
We either use the library file specified by the `HCCL_SO_PATH`
We either use the library file specified by the `FASTVIDEO_HCCL_SO_PATH`
environment variable, or we find the library file brought by PyTorch.
After importing `torch`, `libhccl.so` can be
found by `ctypes` automatically.
"""
so_file = envs.HCCL_SO_PATH.get()
so_file = envs.FASTVIDEO_HCCL_SO_PATH.get()
# manually load the nccl library
if so_file:
logger.info("Found hccl from environment variable HCCL_SO_PATH=%s", so_file)
logger.info("Found hccl from environment variable FASTVIDEO_HCCL_SO_PATH=%s", so_file)
else:
if torch.version.cann is not None: # codespell:ignore cann
so_file = "libhccl.so"
@@ -1240,9 +1240,9 @@ def xpu_is_initialized() -> bool:
def force_spawn() -> None:
if os.environ.get("FASTVIDEO_WORKER_MULTIPROC_METHOD") == "fork":
if envs.FASTVIDEO_WORKER_MULTIPROC_METHOD.get() == "fork":
logger.warning("We must use the `spawn` multiprocessing start method.")
os.environ["FASTVIDEO_WORKER_MULTIPROC_METHOD"] = "spawn"
envs.FASTVIDEO_WORKER_MULTIPROC_METHOD.set("spawn")
def get_mp_context() -> BaseContext:
fastvideo/worker/gpu_worker.py
+5 -6
View File
@@ -1,5 +1,4 @@
# SPDX-License-Identifier: Apache-2.0
import os
from typing import Any, cast
import torch
@@ -50,9 +49,9 @@ class Worker:
# this behavior.
# Related issue:
# https://discuss.pytorch.org/t/cuda-allocation-lifetime-for-inputs-to-distributed-all-reduce/191573
os.environ["TORCH_NCCL_AVOID_RECORD_STREAMS"] = "1"
envs.set_external("TORCH_NCCL_AVOID_RECORD_STREAMS", "1")
# This env var set by Ray causes exceptions with graph building.
os.environ.pop("NCCL_ASYNC_ERROR_HANDLING", None)
envs.unset_external("NCCL_ASYNC_ERROR_HANDLING")
# Set environment variables BEFORE calling get_local_torch_device()
# so that each worker uses the correct device
@@ -60,9 +59,9 @@ class Worker:
# Ray deliberately excludes LOCAL_RANK from the copied driver
# environment and exposes all GPUs assigned to the node, so leaving an
# inherited or missing value here would bind every Ray actor to cuda:0.
os.environ["LOCAL_RANK"] = str(self.local_rank)
os.environ["RANK"] = str(self.rank)
os.environ["WORLD_SIZE"] = str(self.fastvideo_args.num_gpus)
envs.set_external("LOCAL_RANK", str(self.local_rank))
envs.set_external("RANK", str(self.rank))
envs.set_external("WORLD_SIZE", str(self.fastvideo_args.num_gpus))
# Platform-agnostic device initialization
self.device = get_local_torch_device()
fastvideo/worker/ray_distributed_executor.py
+2 -1
View File
@@ -94,6 +94,7 @@ class RayDistributedExecutor(Executor):
"NCCL_NVLS_ENABLE",
"NCCL_DEBUG",
"NCCL_DEBUG_SUBSYS",
"LD_LIBRARY_PATH",
}
def _init_executor(self) -> None:
@@ -103,7 +104,7 @@ class RayDistributedExecutor(Executor):
# Disable Ray usage stats collection.
ray_usage = os.environ.get("RAY_USAGE_STATS_ENABLED", "0")
if ray_usage != "1":
os.environ["RAY_USAGE_STATS_ENABLED"] = "0"
envs.set_external("RAY_USAGE_STATS_ENABLED", "0")
self._init_workers_ray(placement_group)
fastvideo/worker/ray_env.py
+25 -21
View File
@@ -8,24 +8,20 @@ from fastvideo.logger import init_logger
logger = init_logger(__name__)
CONFIG_HOME = envs.FASTVIDEO_CONFIG_ROOT.get()
# This file contains a list of env vars that should not be copied
# from the driver to the Ray workers.
RAY_NON_CARRY_OVER_ENV_VARS_FILE = os.path.join(CONFIG_HOME, "ray_non_carry_over_env_vars.json")
try:
if os.path.exists(RAY_NON_CARRY_OVER_ENV_VARS_FILE):
with open(RAY_NON_CARRY_OVER_ENV_VARS_FILE) as f:
RAY_NON_CARRY_OVER_ENV_VARS = set(json.load(f))
else:
RAY_NON_CARRY_OVER_ENV_VARS = set()
except json.JSONDecodeError:
logger.warning(
"Failed to parse %s. Using an empty set for non-carry-over env vars.",
RAY_NON_CARRY_OVER_ENV_VARS_FILE,
)
RAY_NON_CARRY_OVER_ENV_VARS = set()
def _load_non_carry_over_env_vars(non_carry_over_file: str) -> set[str]:
"""Read the env vars that must not be copied from the driver to the Ray workers."""
try:
if os.path.exists(non_carry_over_file):
with open(non_carry_over_file) as f:
return set(json.load(f))
return set()
except json.JSONDecodeError:
logger.warning(
"Failed to parse %s. Using an empty set for non-carry-over env vars.",
non_carry_over_file,
)
return set()
def get_env_vars_to_copy(
@@ -52,18 +48,26 @@ def get_env_vars_to_copy(
"""
exclude_vars = exclude_vars or set()
additional_vars = additional_vars or set()
# This file contains a list of env vars that should not be copied
# from the driver to the Ray workers.
non_carry_over_file = os.path.join(envs.FASTVIDEO_CONFIG_ROOT.get(), "ray_non_carry_over_env_vars.json")
non_carry_over_vars = _load_non_carry_over_env_vars(non_carry_over_file)
registered_vars = {
name
for field in envs.environment_variables.values()
for name in (field.name, *field.deprecated_names)
}
env_vars_to_copy = {
v
for v in set(envs.environment_variables).union(additional_vars)
if v not in exclude_vars and v not in RAY_NON_CARRY_OVER_ENV_VARS
for v in registered_vars.union(additional_vars) if v not in exclude_vars and v not in non_carry_over_vars
}
to_destination = " to " + destination if destination is not None else ""
logger.info(
"RAY_NON_CARRY_OVER_ENV_VARS from config: %s",
RAY_NON_CARRY_OVER_ENV_VARS,
non_carry_over_vars,
)
logger.info(
"Copying the following environment variables%s: %s",
@@ -72,7 +76,7 @@ def get_env_vars_to_copy(
)
logger.info(
"If certain env vars should NOT be copied, add them to %s file",
RAY_NON_CARRY_OVER_ENV_VARS_FILE,
non_carry_over_file,
)
return env_vars_to_copy
tests/local_tests/ltx2/README.md
+1 -1
View File
@@ -16,7 +16,7 @@ skip or fail. Skipped in CI; run locally on a single GPU.
| Official commit/version | `<TODO>` |
| HF weights | `Lightricks/LTX-2`, `FastVideo/LTX2-base`, `FastVideo/LTX2-Distilled-Diffusers` |
| HF revision | `<TODO>` |
| Local weights dir | `<TODO>` (env: `LTX2_DIFFUSERS_PATH`, `LTX2_FASTVIDEO_GEMMA_LOG`, etc.) |
| Local weights dir | `<TODO>` (env: `LTX2_DIFFUSERS_PATH`, `FASTVIDEO_LTX2_GEMMA_LOG`, etc.) |
| Source layout | `<TODO: diffusers / raw_official / mixed>` |
| Needs conversion | `<TODO>` |
tests/local_tests/ltx2/test_ltx2_gemma_parity.py
+1 -1
View File
@@ -21,7 +21,7 @@ if ltx_core_path.exists() and str(ltx_core_path) not in sys.path:
def _init_log_paths() -> tuple[Path, Path]:
base_dir = Path(os.getenv("LTX2_DEBUG_DIR", "ltx2_debug"))
fastvideo_log = Path(os.getenv("LTX2_FASTVIDEO_GEMMA_LOG", base_dir / "fastvideo_gemma.log"))
fastvideo_log = Path(os.getenv("FASTVIDEO_LTX2_GEMMA_LOG", base_dir / "fastvideo_gemma.log"))
reference_log = Path(os.getenv("LTX2_REFERENCE_GEMMA_LOG", base_dir / "reference_gemma.log"))
fastvideo_log.parent.mkdir(parents=True, exist_ok=True)
reference_log.parent.mkdir(parents=True, exist_ok=True)