Built site for gh-pages
This commit is contained in:
Quarto GHA Workflow Runner
187
docs/config.html
187
docs/config.html
@@ -698,10 +698,10 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
|
||||
<span id="cb1-232"><a href="#cb1-232" aria-hidden="true" tabindex="-1"></a><span class="co"># grpo</span></span>
|
||||
<span id="cb1-233"><a href="#cb1-233" aria-hidden="true" tabindex="-1"></a><span class="fu">trl</span><span class="kw">:</span></span>
|
||||
<span id="cb1-234"><a href="#cb1-234" aria-hidden="true" tabindex="-1"></a><span class="at"> </span><span class="fu">use_vllm</span><span class="kw">:</span><span class="co"> # Optional[bool]. Whether to use VLLM for RL training.</span></span>
|
||||
<span id="cb1-235"><a href="#cb1-235" aria-hidden="true" tabindex="-1"></a><span class="at"> </span><span class="fu">vllm_device</span><span class="kw">:</span><span class="co"> # Optional[str]. Device to use for VLLM.</span></span>
|
||||
<span id="cb1-236"><a href="#cb1-236" aria-hidden="true" tabindex="-1"></a><span class="at"> </span><span class="fu">vllm_gpu_memory_utilization</span><span class="kw">:</span><span class="co"> # Optional[float]. GPU memory utilization for VLLM.</span></span>
|
||||
<span id="cb1-237"><a href="#cb1-237" aria-hidden="true" tabindex="-1"></a><span class="at"> </span><span class="fu">vllm_max_model_len</span><span class="kw">:</span><span class="co"> # Optional[int]. Maximum length of the model for VLLM.</span></span>
|
||||
<span id="cb1-238"><a href="#cb1-238" aria-hidden="true" tabindex="-1"></a><span class="at"> </span><span class="fu">vllm_dtype</span><span class="kw">:</span><span class="co"> # Optional[str]. Data type for VLLM.</span></span>
|
||||
<span id="cb1-235"><a href="#cb1-235" aria-hidden="true" tabindex="-1"></a><span class="at"> </span><span class="fu">vllm_server_host</span><span class="kw">:</span><span class="co"> # Optional[str]. Host of the vLLM server to connect to.</span></span>
|
||||
<span id="cb1-236"><a href="#cb1-236" aria-hidden="true" tabindex="-1"></a><span class="at"> </span><span class="fu">vllm_server_port</span><span class="kw">:</span><span class="co"> # Optional[int]. Port of the vLLM server to connect to.</span></span>
|
||||
<span id="cb1-237"><a href="#cb1-237" aria-hidden="true" tabindex="-1"></a><span class="at"> </span><span class="fu">vllm_server_timeout</span><span class="kw">:</span><span class="co"> # Optional[int]. Total timeout (in seconds) to wait for the vLLM server to respond.</span></span>
|
||||
<span id="cb1-238"><a href="#cb1-238" aria-hidden="true" tabindex="-1"></a><span class="at"> </span><span class="fu">vllm_guided_decoding_regex</span><span class="kw">:</span><span class="co"> # Optional[str]. Regex for vLLM guided decoding.</span></span>
|
||||
<span id="cb1-239"><a href="#cb1-239" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-240"><a href="#cb1-240" aria-hidden="true" tabindex="-1"></a><span class="at"> </span><span class="fu">beta</span><span class="kw">:</span><span class="co"> # Optional[float]. Beta parameter for the RL training. Same as `rl_beta`. Use</span></span>
|
||||
<span id="cb1-241"><a href="#cb1-241" aria-hidden="true" tabindex="-1"></a><span class="at"> </span><span class="fu">max_completion_length</span><span class="kw">:</span><span class="co"> # Optional[int]. Maximum length of the completion for RL training.</span></span>
|
||||
@@ -1047,95 +1047,100 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
|
||||
<span id="cb1-581"><a href="#cb1-581" aria-hidden="true" tabindex="-1"></a><span class="co"># currently only supported on Llama and Mistral</span></span>
|
||||
<span id="cb1-582"><a href="#cb1-582" aria-hidden="true" tabindex="-1"></a><span class="fu">neftune_noise_alpha</span><span class="kw">:</span></span>
|
||||
<span id="cb1-583"><a href="#cb1-583" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-584"><a href="#cb1-584" aria-hidden="true" tabindex="-1"></a><span class="co"># Whether to bettertransformers</span></span>
|
||||
<span id="cb1-584"><a href="#cb1-584" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[bool]. Whether to bettertransformers</span></span>
|
||||
<span id="cb1-585"><a href="#cb1-585" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_optimum</span><span class="kw">:</span></span>
|
||||
<span id="cb1-586"><a href="#cb1-586" aria-hidden="true" tabindex="-1"></a><span class="co"># Whether to use xformers attention patch https://github.com/facebookresearch/xformers:</span></span>
|
||||
<span id="cb1-587"><a href="#cb1-587" aria-hidden="true" tabindex="-1"></a><span class="fu">xformers_attention</span><span class="kw">:</span></span>
|
||||
<span id="cb1-588"><a href="#cb1-588" aria-hidden="true" tabindex="-1"></a><span class="co"># Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention:</span></span>
|
||||
<span id="cb1-589"><a href="#cb1-589" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attention</span><span class="kw">:</span></span>
|
||||
<span id="cb1-590"><a href="#cb1-590" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attn_cross_entropy</span><span class="kw">:</span><span class="co"> # Whether to use flash-attention cross entropy implementation - advanced use only</span></span>
|
||||
<span id="cb1-591"><a href="#cb1-591" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attn_rms_norm</span><span class="kw">:</span><span class="co"> # Whether to use flash-attention rms norm implementation - advanced use only</span></span>
|
||||
<span id="cb1-592"><a href="#cb1-592" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attn_fuse_qkv</span><span class="kw">:</span><span class="co"> # Whether to fuse QKV into a single operation</span></span>
|
||||
<span id="cb1-593"><a href="#cb1-593" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attn_fuse_mlp</span><span class="kw">:</span><span class="co"> # Whether to fuse part of the MLP into a single operation</span></span>
|
||||
<span id="cb1-594"><a href="#cb1-594" aria-hidden="true" tabindex="-1"></a><span class="co"># Whether to use scaled-dot-product attention</span></span>
|
||||
<span id="cb1-595"><a href="#cb1-595" aria-hidden="true" tabindex="-1"></a><span class="co"># https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html</span></span>
|
||||
<span id="cb1-596"><a href="#cb1-596" aria-hidden="true" tabindex="-1"></a><span class="fu">sdp_attention</span><span class="kw">:</span></span>
|
||||
<span id="cb1-597"><a href="#cb1-597" aria-hidden="true" tabindex="-1"></a><span class="co"># Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf</span></span>
|
||||
<span id="cb1-598"><a href="#cb1-598" aria-hidden="true" tabindex="-1"></a><span class="fu">s2_attention</span><span class="kw">:</span></span>
|
||||
<span id="cb1-599"><a href="#cb1-599" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[bool]. Whether to use low_cpu_mem_usage</span></span>
|
||||
<span id="cb1-600"><a href="#cb1-600" aria-hidden="true" tabindex="-1"></a><span class="fu">low_cpu_mem_usage</span><span class="kw">:</span></span>
|
||||
<span id="cb1-601"><a href="#cb1-601" aria-hidden="true" tabindex="-1"></a><span class="co"># Resume from a specific checkpoint dir</span></span>
|
||||
<span id="cb1-602"><a href="#cb1-602" aria-hidden="true" tabindex="-1"></a><span class="fu">resume_from_checkpoint</span><span class="kw">:</span></span>
|
||||
<span id="cb1-603"><a href="#cb1-603" aria-hidden="true" tabindex="-1"></a><span class="co"># If resume_from_checkpoint isn't set and you simply want it to start where it left off.</span></span>
|
||||
<span id="cb1-604"><a href="#cb1-604" aria-hidden="true" tabindex="-1"></a><span class="co"># Be careful with this being turned on between different models.</span></span>
|
||||
<span id="cb1-605"><a href="#cb1-605" aria-hidden="true" tabindex="-1"></a><span class="fu">auto_resume_from_checkpoints</span><span class="kw">:</span><span class="at"> </span><span class="ch">false</span></span>
|
||||
<span id="cb1-606"><a href="#cb1-606" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-607"><a href="#cb1-607" aria-hidden="true" tabindex="-1"></a><span class="co">## Multimodal section</span></span>
|
||||
<span id="cb1-608"><a href="#cb1-608" aria-hidden="true" tabindex="-1"></a><span class="co"># int | tuple[int, int] | None . Size to resize images to, width x height.</span></span>
|
||||
<span id="cb1-609"><a href="#cb1-609" aria-hidden="true" tabindex="-1"></a><span class="co"># Will read from model/processor config if not set.</span></span>
|
||||
<span id="cb1-610"><a href="#cb1-610" aria-hidden="true" tabindex="-1"></a><span class="fu">image_size</span><span class="kw">:</span></span>
|
||||
<span id="cb1-611"><a href="#cb1-611" aria-hidden="true" tabindex="-1"></a><span class="co"># str. Algorithm to use for image resizing. "bilinear", "bicubic", "lanczos". Default is "bilinear".</span></span>
|
||||
<span id="cb1-612"><a href="#cb1-612" aria-hidden="true" tabindex="-1"></a><span class="fu">image_resize_algorithm</span><span class="kw">:</span><span class="at"> </span><span class="st">'bilinear'</span></span>
|
||||
<span id="cb1-613"><a href="#cb1-613" aria-hidden="true" tabindex="-1"></a><span class="co">## End of multimodal section</span></span>
|
||||
<span id="cb1-614"><a href="#cb1-614" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-615"><a href="#cb1-615" aria-hidden="true" tabindex="-1"></a><span class="co"># Don't mess with this, it's here for accelerate and torchrun</span></span>
|
||||
<span id="cb1-616"><a href="#cb1-616" aria-hidden="true" tabindex="-1"></a><span class="fu">local_rank</span><span class="kw">:</span></span>
|
||||
<span id="cb1-617"><a href="#cb1-617" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-618"><a href="#cb1-618" aria-hidden="true" tabindex="-1"></a><span class="co"># Add or change special tokens.</span></span>
|
||||
<span id="cb1-619"><a href="#cb1-619" aria-hidden="true" tabindex="-1"></a><span class="co"># If you add tokens here, you don't need to add them to the `tokens` list.</span></span>
|
||||
<span id="cb1-620"><a href="#cb1-620" aria-hidden="true" tabindex="-1"></a><span class="fu">special_tokens</span><span class="kw">:</span></span>
|
||||
<span id="cb1-621"><a href="#cb1-621" aria-hidden="true" tabindex="-1"></a><span class="co"> # bos_token: "<s>"</span></span>
|
||||
<span id="cb1-622"><a href="#cb1-622" aria-hidden="true" tabindex="-1"></a><span class="co"> # eos_token: "</s>"</span></span>
|
||||
<span id="cb1-623"><a href="#cb1-623" aria-hidden="true" tabindex="-1"></a><span class="co"> # unk_token: "<unk>"</span></span>
|
||||
<span id="cb1-624"><a href="#cb1-624" aria-hidden="true" tabindex="-1"></a><span class="co"> # pad_token: "[PAD]"</span></span>
|
||||
<span id="cb1-625"><a href="#cb1-625" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-626"><a href="#cb1-626" aria-hidden="true" tabindex="-1"></a><span class="co"># Add extra tokens.</span></span>
|
||||
<span id="cb1-627"><a href="#cb1-627" aria-hidden="true" tabindex="-1"></a><span class="fu">tokens</span><span class="kw">:</span></span>
|
||||
<span id="cb1-628"><a href="#cb1-628" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-629"><a href="#cb1-629" aria-hidden="true" tabindex="-1"></a><span class="co"># Mapping token_id to new_token_string to override reserved added_tokens in the tokenizer.</span></span>
|
||||
<span id="cb1-630"><a href="#cb1-630" aria-hidden="true" tabindex="-1"></a><span class="co"># Only works for tokens that are not part of the base vocab (aka are added_tokens).</span></span>
|
||||
<span id="cb1-631"><a href="#cb1-631" aria-hidden="true" tabindex="-1"></a><span class="co"># Can be checked if they exist in tokenizer.json added_tokens.</span></span>
|
||||
<span id="cb1-632"><a href="#cb1-632" aria-hidden="true" tabindex="-1"></a><span class="fu">added_tokens_overrides</span><span class="kw">:</span><span class="co"> # Dict[int, str]</span></span>
|
||||
<span id="cb1-633"><a href="#cb1-633" aria-hidden="true" tabindex="-1"></a><span class="co"># 128041: "<|im_start|>"</span></span>
|
||||
<span id="cb1-634"><a href="#cb1-634" aria-hidden="true" tabindex="-1"></a><span class="co"># 128042: "<|im_end|>"</span></span>
|
||||
<span id="cb1-635"><a href="#cb1-635" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-636"><a href="#cb1-636" aria-hidden="true" tabindex="-1"></a><span class="co"># FSDP</span></span>
|
||||
<span id="cb1-637"><a href="#cb1-637" aria-hidden="true" tabindex="-1"></a><span class="fu">fsdp</span><span class="kw">:</span></span>
|
||||
<span id="cb1-638"><a href="#cb1-638" aria-hidden="true" tabindex="-1"></a><span class="fu">fsdp_config</span><span class="kw">:</span></span>
|
||||
<span id="cb1-639"><a href="#cb1-639" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-640"><a href="#cb1-640" aria-hidden="true" tabindex="-1"></a><span class="co"># Deepspeed config path. e.g., deepspeed_configs/zero3.json</span></span>
|
||||
<span id="cb1-641"><a href="#cb1-641" aria-hidden="true" tabindex="-1"></a><span class="fu">deepspeed</span><span class="kw">:</span></span>
|
||||
<span id="cb1-642"><a href="#cb1-642" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-643"><a href="#cb1-643" aria-hidden="true" tabindex="-1"></a><span class="co"># Advanced DDP Arguments</span></span>
|
||||
<span id="cb1-644"><a href="#cb1-644" aria-hidden="true" tabindex="-1"></a><span class="fu">ddp_timeout</span><span class="kw">:</span></span>
|
||||
<span id="cb1-645"><a href="#cb1-645" aria-hidden="true" tabindex="-1"></a><span class="fu">ddp_bucket_cap_mb</span><span class="kw">:</span></span>
|
||||
<span id="cb1-646"><a href="#cb1-646" aria-hidden="true" tabindex="-1"></a><span class="fu">ddp_broadcast_buffers</span><span class="kw">:</span></span>
|
||||
<span id="cb1-586"><a href="#cb1-586" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-587"><a href="#cb1-587" aria-hidden="true" tabindex="-1"></a><span class="co"># Note: Only one of the following attention patches can be used at a time.</span></span>
|
||||
<span id="cb1-588"><a href="#cb1-588" aria-hidden="true" tabindex="-1"></a><span class="co"># For example, if you set `xformers_attention` to `true`, do not set `flash_attention` to `true`.</span></span>
|
||||
<span id="cb1-589"><a href="#cb1-589" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-590"><a href="#cb1-590" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[bool]. Whether to use xformers attention patch https://github.com/facebookresearch/xformers:</span></span>
|
||||
<span id="cb1-591"><a href="#cb1-591" aria-hidden="true" tabindex="-1"></a><span class="fu">xformers_attention</span><span class="kw">:</span></span>
|
||||
<span id="cb1-592"><a href="#cb1-592" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[bool]. Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention:</span></span>
|
||||
<span id="cb1-593"><a href="#cb1-593" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attention</span><span class="kw">:</span></span>
|
||||
<span id="cb1-594"><a href="#cb1-594" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attn_cross_entropy</span><span class="kw">:</span><span class="co"> # Optional[bool]. Whether to use flash-attention cross entropy implementation - advanced use only</span></span>
|
||||
<span id="cb1-595"><a href="#cb1-595" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attn_rms_norm</span><span class="kw">:</span><span class="co"> # Optional[bool]. Whether to use flash-attention rms norm implementation - advanced use only</span></span>
|
||||
<span id="cb1-596"><a href="#cb1-596" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attn_fuse_qkv</span><span class="kw">:</span><span class="co"> # Optional[bool]. Whether to fuse QKV into a single operation</span></span>
|
||||
<span id="cb1-597"><a href="#cb1-597" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attn_fuse_mlp</span><span class="kw">:</span><span class="co"> # Optional[bool]. Whether to fuse part of the MLP into a single operation</span></span>
|
||||
<span id="cb1-598"><a href="#cb1-598" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[bool]. Whether to use scaled-dot-product attention</span></span>
|
||||
<span id="cb1-599"><a href="#cb1-599" aria-hidden="true" tabindex="-1"></a><span class="co"># https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html</span></span>
|
||||
<span id="cb1-600"><a href="#cb1-600" aria-hidden="true" tabindex="-1"></a><span class="fu">sdp_attention</span><span class="kw">:</span></span>
|
||||
<span id="cb1-601"><a href="#cb1-601" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[bool]. Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf</span></span>
|
||||
<span id="cb1-602"><a href="#cb1-602" aria-hidden="true" tabindex="-1"></a><span class="fu">s2_attention</span><span class="kw">:</span></span>
|
||||
<span id="cb1-603"><a href="#cb1-603" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-604"><a href="#cb1-604" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[bool]. Whether to use low_cpu_mem_usage</span></span>
|
||||
<span id="cb1-605"><a href="#cb1-605" aria-hidden="true" tabindex="-1"></a><span class="fu">low_cpu_mem_usage</span><span class="kw">:</span></span>
|
||||
<span id="cb1-606"><a href="#cb1-606" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[str]. Resume from a specific checkpoint dir</span></span>
|
||||
<span id="cb1-607"><a href="#cb1-607" aria-hidden="true" tabindex="-1"></a><span class="fu">resume_from_checkpoint</span><span class="kw">:</span></span>
|
||||
<span id="cb1-608"><a href="#cb1-608" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[bool]. If resume_from_checkpoint isn't set and you simply want it to start where it left off.</span></span>
|
||||
<span id="cb1-609"><a href="#cb1-609" aria-hidden="true" tabindex="-1"></a><span class="co"># Be careful with this being turned on between different models.</span></span>
|
||||
<span id="cb1-610"><a href="#cb1-610" aria-hidden="true" tabindex="-1"></a><span class="fu">auto_resume_from_checkpoints</span><span class="kw">:</span><span class="at"> </span><span class="ch">false</span></span>
|
||||
<span id="cb1-611"><a href="#cb1-611" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-612"><a href="#cb1-612" aria-hidden="true" tabindex="-1"></a><span class="co">## Multimodal section</span></span>
|
||||
<span id="cb1-613"><a href="#cb1-613" aria-hidden="true" tabindex="-1"></a><span class="co"># int | tuple[int, int] | None . Size to resize images to, width x height.</span></span>
|
||||
<span id="cb1-614"><a href="#cb1-614" aria-hidden="true" tabindex="-1"></a><span class="co"># Will read from model/processor config if not set.</span></span>
|
||||
<span id="cb1-615"><a href="#cb1-615" aria-hidden="true" tabindex="-1"></a><span class="fu">image_size</span><span class="kw">:</span></span>
|
||||
<span id="cb1-616"><a href="#cb1-616" aria-hidden="true" tabindex="-1"></a><span class="co"># str. Algorithm to use for image resizing. "bilinear", "bicubic", "lanczos". Default is "bilinear".</span></span>
|
||||
<span id="cb1-617"><a href="#cb1-617" aria-hidden="true" tabindex="-1"></a><span class="fu">image_resize_algorithm</span><span class="kw">:</span><span class="at"> </span><span class="st">'bilinear'</span></span>
|
||||
<span id="cb1-618"><a href="#cb1-618" aria-hidden="true" tabindex="-1"></a><span class="co">## End of multimodal section</span></span>
|
||||
<span id="cb1-619"><a href="#cb1-619" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-620"><a href="#cb1-620" aria-hidden="true" tabindex="-1"></a><span class="co"># Don't mess with this, it's here for accelerate and torchrun</span></span>
|
||||
<span id="cb1-621"><a href="#cb1-621" aria-hidden="true" tabindex="-1"></a><span class="fu">local_rank</span><span class="kw">:</span></span>
|
||||
<span id="cb1-622"><a href="#cb1-622" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-623"><a href="#cb1-623" aria-hidden="true" tabindex="-1"></a><span class="co"># Add or change special tokens.</span></span>
|
||||
<span id="cb1-624"><a href="#cb1-624" aria-hidden="true" tabindex="-1"></a><span class="co"># If you add tokens here, you don't need to add them to the `tokens` list.</span></span>
|
||||
<span id="cb1-625"><a href="#cb1-625" aria-hidden="true" tabindex="-1"></a><span class="fu">special_tokens</span><span class="kw">:</span></span>
|
||||
<span id="cb1-626"><a href="#cb1-626" aria-hidden="true" tabindex="-1"></a><span class="co"> # bos_token: "<s>"</span></span>
|
||||
<span id="cb1-627"><a href="#cb1-627" aria-hidden="true" tabindex="-1"></a><span class="co"> # eos_token: "</s>"</span></span>
|
||||
<span id="cb1-628"><a href="#cb1-628" aria-hidden="true" tabindex="-1"></a><span class="co"> # unk_token: "<unk>"</span></span>
|
||||
<span id="cb1-629"><a href="#cb1-629" aria-hidden="true" tabindex="-1"></a><span class="co"> # pad_token: "[PAD]"</span></span>
|
||||
<span id="cb1-630"><a href="#cb1-630" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-631"><a href="#cb1-631" aria-hidden="true" tabindex="-1"></a><span class="co"># Add extra tokens.</span></span>
|
||||
<span id="cb1-632"><a href="#cb1-632" aria-hidden="true" tabindex="-1"></a><span class="fu">tokens</span><span class="kw">:</span></span>
|
||||
<span id="cb1-633"><a href="#cb1-633" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-634"><a href="#cb1-634" aria-hidden="true" tabindex="-1"></a><span class="co"># Mapping token_id to new_token_string to override reserved added_tokens in the tokenizer.</span></span>
|
||||
<span id="cb1-635"><a href="#cb1-635" aria-hidden="true" tabindex="-1"></a><span class="co"># Only works for tokens that are not part of the base vocab (aka are added_tokens).</span></span>
|
||||
<span id="cb1-636"><a href="#cb1-636" aria-hidden="true" tabindex="-1"></a><span class="co"># Can be checked if they exist in tokenizer.json added_tokens.</span></span>
|
||||
<span id="cb1-637"><a href="#cb1-637" aria-hidden="true" tabindex="-1"></a><span class="fu">added_tokens_overrides</span><span class="kw">:</span><span class="co"> # Dict[int, str]</span></span>
|
||||
<span id="cb1-638"><a href="#cb1-638" aria-hidden="true" tabindex="-1"></a><span class="co"># 128041: "<|im_start|>"</span></span>
|
||||
<span id="cb1-639"><a href="#cb1-639" aria-hidden="true" tabindex="-1"></a><span class="co"># 128042: "<|im_end|>"</span></span>
|
||||
<span id="cb1-640"><a href="#cb1-640" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-641"><a href="#cb1-641" aria-hidden="true" tabindex="-1"></a><span class="co"># FSDP</span></span>
|
||||
<span id="cb1-642"><a href="#cb1-642" aria-hidden="true" tabindex="-1"></a><span class="fu">fsdp</span><span class="kw">:</span></span>
|
||||
<span id="cb1-643"><a href="#cb1-643" aria-hidden="true" tabindex="-1"></a><span class="fu">fsdp_config</span><span class="kw">:</span></span>
|
||||
<span id="cb1-644"><a href="#cb1-644" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-645"><a href="#cb1-645" aria-hidden="true" tabindex="-1"></a><span class="co"># Deepspeed config path. e.g., deepspeed_configs/zero3.json</span></span>
|
||||
<span id="cb1-646"><a href="#cb1-646" aria-hidden="true" tabindex="-1"></a><span class="fu">deepspeed</span><span class="kw">:</span></span>
|
||||
<span id="cb1-647"><a href="#cb1-647" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-648"><a href="#cb1-648" aria-hidden="true" tabindex="-1"></a><span class="co"># Sequence parallelism</span></span>
|
||||
<span id="cb1-649"><a href="#cb1-649" aria-hidden="true" tabindex="-1"></a><span class="co"># Set to a divisor of the number of GPUs available to split sequences into chunks of equal size.</span></span>
|
||||
<span id="cb1-650"><a href="#cb1-650" aria-hidden="true" tabindex="-1"></a><span class="co"># Use in long context training to prevent OOM when sequences cannot fit into a single GPU's VRAM.</span></span>
|
||||
<span id="cb1-651"><a href="#cb1-651" aria-hidden="true" tabindex="-1"></a><span class="co"># E.g., if 4 GPUs are available, set this value to 2 to split each sequence into two equal-sized</span></span>
|
||||
<span id="cb1-652"><a href="#cb1-652" aria-hidden="true" tabindex="-1"></a><span class="co"># subsequences, or set to 4 to split into four equal-sized subsequences.</span></span>
|
||||
<span id="cb1-653"><a href="#cb1-653" aria-hidden="true" tabindex="-1"></a><span class="co"># See https://axolotl-ai-cloud.github.io/axolotl/docs/sequence_parallelism.html for more details.</span></span>
|
||||
<span id="cb1-654"><a href="#cb1-654" aria-hidden="true" tabindex="-1"></a><span class="fu">sequence_parallel_degree</span><span class="kw">:</span></span>
|
||||
<span id="cb1-655"><a href="#cb1-655" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional; strides across the key dimension. Larger values use more memory but should make training faster.</span></span>
|
||||
<span id="cb1-656"><a href="#cb1-656" aria-hidden="true" tabindex="-1"></a><span class="co"># Must evenly divide the number of KV heads in your model.</span></span>
|
||||
<span id="cb1-657"><a href="#cb1-657" aria-hidden="true" tabindex="-1"></a><span class="fu">heads_k_stride</span><span class="kw">:</span><span class="at"> </span><span class="dv">1</span></span>
|
||||
<span id="cb1-658"><a href="#cb1-658" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-659"><a href="#cb1-659" aria-hidden="true" tabindex="-1"></a><span class="co"># Path to torch distx for optim 'adamw_anyprecision'</span></span>
|
||||
<span id="cb1-660"><a href="#cb1-660" aria-hidden="true" tabindex="-1"></a><span class="fu">torchdistx_path</span><span class="kw">:</span></span>
|
||||
<span id="cb1-661"><a href="#cb1-661" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-662"><a href="#cb1-662" aria-hidden="true" tabindex="-1"></a><span class="co"># Set to HF dataset for type: 'completion' for streaming instead of pre-tokenize</span></span>
|
||||
<span id="cb1-663"><a href="#cb1-663" aria-hidden="true" tabindex="-1"></a><span class="fu">pretraining_dataset</span><span class="kw">:</span></span>
|
||||
<span id="cb1-664"><a href="#cb1-664" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-665"><a href="#cb1-665" aria-hidden="true" tabindex="-1"></a><span class="co"># Debug mode</span></span>
|
||||
<span id="cb1-666"><a href="#cb1-666" aria-hidden="true" tabindex="-1"></a><span class="fu">debug</span><span class="kw">:</span></span>
|
||||
<span id="cb1-667"><a href="#cb1-667" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-668"><a href="#cb1-668" aria-hidden="true" tabindex="-1"></a><span class="co"># Seed</span></span>
|
||||
<span id="cb1-669"><a href="#cb1-669" aria-hidden="true" tabindex="-1"></a><span class="fu">seed</span><span class="kw">:</span></span>
|
||||
<span id="cb1-670"><a href="#cb1-670" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-671"><a href="#cb1-671" aria-hidden="true" tabindex="-1"></a><span class="co"># Allow overwrite yml config using from cli</span></span>
|
||||
<span id="cb1-672"><a href="#cb1-672" aria-hidden="true" tabindex="-1"></a><span class="fu">strict</span><span class="kw">:</span></span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
|
||||
<span id="cb1-648"><a href="#cb1-648" aria-hidden="true" tabindex="-1"></a><span class="co"># Advanced DDP Arguments</span></span>
|
||||
<span id="cb1-649"><a href="#cb1-649" aria-hidden="true" tabindex="-1"></a><span class="fu">ddp_timeout</span><span class="kw">:</span></span>
|
||||
<span id="cb1-650"><a href="#cb1-650" aria-hidden="true" tabindex="-1"></a><span class="fu">ddp_bucket_cap_mb</span><span class="kw">:</span></span>
|
||||
<span id="cb1-651"><a href="#cb1-651" aria-hidden="true" tabindex="-1"></a><span class="fu">ddp_broadcast_buffers</span><span class="kw">:</span></span>
|
||||
<span id="cb1-652"><a href="#cb1-652" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-653"><a href="#cb1-653" aria-hidden="true" tabindex="-1"></a><span class="co"># Sequence parallelism</span></span>
|
||||
<span id="cb1-654"><a href="#cb1-654" aria-hidden="true" tabindex="-1"></a><span class="co"># Set to a divisor of the number of GPUs available to split sequences into chunks of equal size.</span></span>
|
||||
<span id="cb1-655"><a href="#cb1-655" aria-hidden="true" tabindex="-1"></a><span class="co"># Use in long context training to prevent OOM when sequences cannot fit into a single GPU's VRAM.</span></span>
|
||||
<span id="cb1-656"><a href="#cb1-656" aria-hidden="true" tabindex="-1"></a><span class="co"># E.g., if 4 GPUs are available, set this value to 2 to split each sequence into two equal-sized</span></span>
|
||||
<span id="cb1-657"><a href="#cb1-657" aria-hidden="true" tabindex="-1"></a><span class="co"># subsequences, or set to 4 to split into four equal-sized subsequences.</span></span>
|
||||
<span id="cb1-658"><a href="#cb1-658" aria-hidden="true" tabindex="-1"></a><span class="co"># See https://axolotl-ai-cloud.github.io/axolotl/docs/sequence_parallelism.html for more details.</span></span>
|
||||
<span id="cb1-659"><a href="#cb1-659" aria-hidden="true" tabindex="-1"></a><span class="fu">sequence_parallel_degree</span><span class="kw">:</span></span>
|
||||
<span id="cb1-660"><a href="#cb1-660" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional; strides across the key dimension. Larger values use more memory but should make training faster.</span></span>
|
||||
<span id="cb1-661"><a href="#cb1-661" aria-hidden="true" tabindex="-1"></a><span class="co"># Must evenly divide the number of KV heads in your model.</span></span>
|
||||
<span id="cb1-662"><a href="#cb1-662" aria-hidden="true" tabindex="-1"></a><span class="fu">heads_k_stride</span><span class="kw">:</span><span class="at"> </span><span class="dv">1</span></span>
|
||||
<span id="cb1-663"><a href="#cb1-663" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-664"><a href="#cb1-664" aria-hidden="true" tabindex="-1"></a><span class="co"># Path to torch distx for optim 'adamw_anyprecision'</span></span>
|
||||
<span id="cb1-665"><a href="#cb1-665" aria-hidden="true" tabindex="-1"></a><span class="fu">torchdistx_path</span><span class="kw">:</span></span>
|
||||
<span id="cb1-666"><a href="#cb1-666" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-667"><a href="#cb1-667" aria-hidden="true" tabindex="-1"></a><span class="co"># Set to HF dataset for type: 'completion' for streaming instead of pre-tokenize</span></span>
|
||||
<span id="cb1-668"><a href="#cb1-668" aria-hidden="true" tabindex="-1"></a><span class="fu">pretraining_dataset</span><span class="kw">:</span></span>
|
||||
<span id="cb1-669"><a href="#cb1-669" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-670"><a href="#cb1-670" aria-hidden="true" tabindex="-1"></a><span class="co"># Debug mode</span></span>
|
||||
<span id="cb1-671"><a href="#cb1-671" aria-hidden="true" tabindex="-1"></a><span class="fu">debug</span><span class="kw">:</span></span>
|
||||
<span id="cb1-672"><a href="#cb1-672" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-673"><a href="#cb1-673" aria-hidden="true" tabindex="-1"></a><span class="co"># Seed</span></span>
|
||||
<span id="cb1-674"><a href="#cb1-674" aria-hidden="true" tabindex="-1"></a><span class="fu">seed</span><span class="kw">:</span></span>
|
||||
<span id="cb1-675"><a href="#cb1-675" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-676"><a href="#cb1-676" aria-hidden="true" tabindex="-1"></a><span class="co"># Allow overwrite yml config using from cli</span></span>
|
||||
<span id="cb1-677"><a href="#cb1-677" aria-hidden="true" tabindex="-1"></a><span class="fu">strict</span><span class="kw">:</span></span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
|
||||
|
||||
|
||||
|
||||
|
||||
Reference in New Issue
Block a user