Remove API for matmul_via_f16 (#1197)

Author: EricLBuehler

mistralrs-core/src/lib.rs

@@ -187,7 +187,6 @@ pub struct MistralRsBuilder {
     no_prefix_cache: Option<bool>,
     prefix_cache_n: Option<usize>,
     disable_eos_stop: Option<bool>,
-    gemm_full_precision_f16: Option<bool>,
     throughput_logging_enabled: Option<()>,
 }
 
@@ -202,7 +201,6 @@ impl MistralRsBuilder {
             no_prefix_cache: None,
             prefix_cache_n: None,
             disable_eos_stop: None,
-            gemm_full_precision_f16: None,
             throughput_logging_enabled: None,
         }
     }
@@ -234,11 +232,6 @@ impl MistralRsBuilder {
         self.disable_eos_stop = Some(disable_eos_stop);
         self
     }
-    /// This setting is only applicable on CUDA. If set to false or not specified, this setting enables f16/bf16 reduced precision matmul for GPUs which support it. If set to true, this setting has no effect.
-    pub fn with_gemm_full_precision_f16(mut self, gemm_full_precision: bool) -> Self {
-        self.gemm_full_precision_f16 = Some(gemm_full_precision);
-        self
-    }
     pub fn with_throughput_logging(mut self) -> Self {
         self.throughput_logging_enabled = Some(());
         self
@@ -249,42 +242,6 @@ impl MistralRsBuilder {
     }
 }
 
-#[cfg(feature = "cuda")]
-fn set_gemm_reduced_precision_f16(device: candle_core::Device) {
-    use mistralrs_quant::INHIBIT_GEMM_F16;
-
-    use candle_core::{DType, Tensor};
-
-    let a = Tensor::zeros((2, 2), DType::BF16, &device).unwrap();
-    candle_core::cuda::set_gemm_reduced_precision_bf16(true);
-    match a.matmul(&a) {
-        Ok(_) => tracing::info!("Enabling GEMM reduced precision in BF16."),
-        Err(e) => {
-            if format!("{e:?}").contains("CUBLAS_STATUS_NOT_SUPPORTED") {
-                tracing::info!("GEMM reduced precision in BF16 not supported.");
-                candle_core::cuda::set_gemm_reduced_precision_bf16(false);
-                INHIBIT_GEMM_F16.store(true, std::sync::atomic::Ordering::Relaxed);
-            }
-        }
-    }
-
-    let a = Tensor::zeros((2, 2), DType::F16, &device).unwrap();
-    candle_core::cuda::set_gemm_reduced_precision_f16(true);
-    match a.matmul(&a) {
-        Ok(_) => tracing::info!("Enabling GEMM reduced precision in F16."),
-        Err(e) => {
-            if format!("{e:?}").contains("CUBLAS_STATUS_NOT_SUPPORTED") {
-                tracing::info!("GEMM reduced precision in F16 not supported.");
-                candle_core::cuda::set_gemm_reduced_precision_f16(false);
-                INHIBIT_GEMM_F16.store(true, std::sync::atomic::Ordering::Relaxed);
-            }
-        }
-    }
-}
-
-#[cfg(not(feature = "cuda"))]
-fn set_gemm_reduced_precision_f16(_device: candle_core::Device) {}
-
 impl Drop for MistralRs {
     fn drop(&mut self) {
         ENGINE_INSTRUCTIONS
@@ -305,19 +262,10 @@ impl MistralRs {
             no_prefix_cache,
             prefix_cache_n,
             disable_eos_stop,
-            gemm_full_precision_f16,
             throughput_logging_enabled,
         } = config;
 
         let category = pipeline.try_lock().unwrap().category();
-        let model_supports_reduced_gemm = match category {
-            ModelCategory::Text => true,
-            ModelCategory::Vision { has_conv2d, .. } => !has_conv2d,
-            ModelCategory::Diffusion => true,
-        };
-        if !gemm_full_precision_f16.unwrap_or(false) && model_supports_reduced_gemm {
-            set_gemm_reduced_precision_f16(get_mut_arcmutex!(pipeline).device());
-        }
         setup_cublas_lt_wrapper(get_mut_arcmutex!(pipeline).device());
 
         let truncate_sequence = truncate_sequence.unwrap_or(false);

mistralrs-server/src/main.rs

@@ -478,9 +478,7 @@ async fn main() -> Result<()> {
         .with_opt_log(args.log)
         .with_truncate_sequence(args.truncate_sequence)
         .with_no_kv_cache(args.no_kv_cache)
-        .with_prefix_cache_n(args.prefix_cache_n)
-        .with_gemm_full_precision_f16(args.cpu)
-        .with_gemm_full_precision_f16(args.cpu); // Required to allow `cuda` build to use `--cpu`, #1056
+        .with_prefix_cache_n(args.prefix_cache_n);
 
     if args.interactive_mode {
         interactive_mode(builder.build(), args.throughput_log).await;

mistralrs/src/anymoe.rs

@@ -112,7 +112,6 @@ impl AnyMoeModelBuilder {
 
         let mut runner = MistralRsBuilder::new(pipeline, scheduler_method)
             .with_no_kv_cache(self.base.no_kv_cache)
-            .with_gemm_full_precision_f16(true)
             .with_no_prefix_cache(self.base.prefix_cache_n.is_none());
 
         if let Some(n) = self.base.prefix_cache_n {

mistralrs/src/diffusion_model.rs

@@ -102,8 +102,7 @@ impl DiffusionModelBuilder {
             method: DefaultSchedulerMethod::Fixed(self.max_num_seqs.try_into()?),
         };
 
-        let runner =
-            MistralRsBuilder::new(pipeline, scheduler_method).with_gemm_full_precision_f16(true);
+        let runner = MistralRsBuilder::new(pipeline, scheduler_method);
 
         Ok(Model::new(runner.build()))
     }

mistralrs/src/gguf.rs

@@ -206,7 +206,6 @@ impl GgufModelBuilder {
 
         let mut runner = MistralRsBuilder::new(pipeline, scheduler_method)
             .with_no_kv_cache(self.no_kv_cache)
-            .with_gemm_full_precision_f16(true)
             .with_no_prefix_cache(self.prefix_cache_n.is_none());
 
         if let Some(n) = self.prefix_cache_n {

mistralrs/src/gguf_lora_model.rs

@@ -79,7 +79,6 @@ impl GgufLoraModelBuilder {
 
         let mut runner = MistralRsBuilder::new(pipeline, scheduler_method)
             .with_no_kv_cache(self.gguf_model.no_kv_cache)
-            .with_gemm_full_precision_f16(true)
             .with_no_prefix_cache(self.gguf_model.prefix_cache_n.is_none());
 
         if let Some(n) = self.gguf_model.prefix_cache_n {

mistralrs/src/gguf_xlora_model.rs

@@ -91,7 +91,6 @@ impl GgufXLoraModelBuilder {
 
         let mut runner = MistralRsBuilder::new(pipeline, scheduler_method)
             .with_no_kv_cache(self.gguf_model.no_kv_cache)
-            .with_gemm_full_precision_f16(true)
             .with_no_prefix_cache(self.gguf_model.prefix_cache_n.is_none());
 
         if let Some(n) = self.gguf_model.prefix_cache_n {

mistralrs/src/lora_model.rs

@@ -85,7 +85,6 @@ impl LoraModelBuilder {
 
         let mut runner = MistralRsBuilder::new(pipeline, scheduler_method)
             .with_no_kv_cache(self.text_model.no_kv_cache)
-            .with_gemm_full_precision_f16(true)
             .with_no_prefix_cache(self.text_model.prefix_cache_n.is_none());
 
         if let Some(n) = self.text_model.prefix_cache_n {

mistralrs/src/speculative.rs

@@ -92,8 +92,7 @@ impl TextSpeculativeBuilder {
             self.speculative_config,
         )?));
 
-        let runner =
-            MistralRsBuilder::new(pipeline, scheduler_method).with_gemm_full_precision_f16(true);
+        let runner = MistralRsBuilder::new(pipeline, scheduler_method);
 
         Ok(Model::new(runner.build()))
     }

mistralrs/src/text_model.rs

@@ -309,7 +309,6 @@ impl TextModelBuilder {
 
         let mut runner = MistralRsBuilder::new(pipeline, scheduler_method)
             .with_no_kv_cache(self.no_kv_cache)
-            .with_gemm_full_precision_f16(true)
             .with_no_prefix_cache(self.prefix_cache_n.is_none());
 
         if let Some(n) = self.prefix_cache_n {